A 1D-CNN with advanced data augmentation for robust speech emotion recognition Authors: Neha Chourasia, Chhattar Singh Lamba, Amit Kumar Gupta Published: 2026 ABSTRACT Speech Emotion Recognition (SER) plays a significant role in human–computer interaction by enabling intelligent systems to recognize and respond to human emotional states. Accurate emotion recognition from speech signals is essential for applications such as mental health assessment, healthcare monitoring, virtual assistants, and human–robot interaction. However, traditional SER approaches often rely on manually engineered acoustic features and suffer from limited generalization due to dataset imbalance, speaker variability, and environmental noise. To address these challenges, this study proposes a Conv1D-based deep learning framework for robust speech emotion recognition using complementary acoustic features and data augmentation techniques. The proposed framework utilizes Mel-Frequency Cepstral Coefficients (MFCCs), Zero-Crossing Rate (ZCR), and Root Mean Square Energy (RMSE) extracted from frame-level speech signals to capture spectral, temporal, and energy-related emotional characteristics. The extracted acoustic descriptors are sequentially concatenated into a unified feature representation and provided as input to the Conv1D architecture for hierarchical temporal feature learning. Experiments were conducted using a combined multi-dataset setup consisting of RAVDESS, SAVEE, CREMA-D, and TESS datasets. The dataset was divided into training, validation, and testing subsets, and data augmentation techniques including noise injection, time shifting, pitch modification, and time stretching were applied exclusively to the training samples to improve robustness and class balance. Experimental evaluation across seven emotional categories demonstrates that the proposed framework achieves a test accuracy of 94.91% with a Macro F1-score of 0.94, showing strong recognition capability across diverse emotional speech conditions. The results indicate that the integration of discriminative acoustic feature extraction, sequential feature representation, and Conv1D-based temporal learning significantly improves speech emotion recognition performance. The proposed framework provides an effective and computationally efficient solution suitable for practical real-world SER applications. Introduction Speech emotion recognition (SER) is an important research area in affective computing and human–computer interaction, aiming to automatically identify human emotional states from speech signals. Emotions play a fundamental role in human communication and are expressed through multiple modalities, including facial expressions, body language, and vocal characteristics such as pitch, tone, and energy variations1. Among these modalities, speech signals provide rich emotional information and are widely used for emotion-aware computing systems. Consequently, SER has gained significant attention in various application domains such as healthcare monitoring, mental health assessment, intelligent virtual assistants, human–robot interaction, and smart environments2– 4. Even for humans, identifying emotions is a challenging and complicated endeavour. Many people claim to be able to do this activity effectively, but they frequently have the chance to identify emotions based on a variety of factors, including body language, facial expressions, and prosody or timbre of speech5. Human-computer interaction (HCI) and artificial emotional intelligence (AEI) has come a long way and research is developing quickly due to technological advancements focuses on how people interact with computers and how well systems react6,7. The goal of research is to make human-machine inter-actions as natural as possible8. In recent decades, speech emotion recognition (SER) has been a popular area of study. Human-machine interaction, education, multimedia content management, entertainment, the automotive sector, text-to-speech conversion, and medical diagnostics are just a few of the application areas where SER is used9. With the rapid advancement of artificial intelligence, machine learning (ML) and deep learning (DL) techniques have significantly improved the capability of speech processing systems. Traditional speech emotion recognition methods rely on handcrafted acoustic features such as pitch, spectral features, and energy-based characteristics, which are subsequently classified using machine learning algorithms including k-Nearest Neighbor (kNN), Support Vector Machines (SVM), Hidden Markov Models (HMM), and Artificial Neural Networks (ANN)10. These techniques have been successfully applied in several speech analysis tasks. Recent studies have demonstrated the effectiveness of ML-based approaches such as optimized genetic algorithm–extreme learning machine models for speech emotion recognition and related speech processing problems11. Furthermore, optimization-based extreme learning machine models have also been successfully applied in spoken language identification and voice-based medical diagnosis tasks, demonstrating the potential of advanced machine learning techniques for speech signal analysis12,13. Similar approaches have been explored in voice pathology detection and speech-based disease diagnosis, highlighting the effectiveness of optimized learning algorithms in speech processing applications14– 16. These studies confirm that machine learning and optimization-based neural models play a crucial role in improving the performance of speech analysis systems. In recent years, deep learning models have further enhanced speech emotion recognition by automatically learning hierarchical feature representations from speech signals. Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and transformer-based architectures have been widely used to model temporal and spectral characteristics of speech signals. These architectures enable the automatic extraction of discriminative emotional features and have significantly improved classification accuracy in SER tasks17. However, despite these advancements, several challenges remain in developing robust speech emotion recognition systems. In recent years, deep learning-based feature extraction approaches have been broadly categorized into supervised and unsupervised learning frameworks. Supervised deep learning methods learn discriminative representations from labeled data and have demonstrated strong performance in speech processing, image reconstruction, and pattern recognition tasks. For example, Yan et al. proposed a multimodal reversible-guidance and cyclical knowledge distillation framework for optimized single-image super-resolution reconstruction, demonstrating the effectiveness of supervised representation learning for complex feature extraction tasks18. In contrast, unsupervised learning approaches aim to extract meaningful latent representations without requiring labeled samples, making them suitable for scenarios involving limited annotations, domain adaptation, and transfer learning. Wang et al. introduced a spatial-channel collaborative multi-scale graph interaction deep transfer learning framework for unsupervised rotating machinery fault diagnosis, highlighting the capability of unsupervised deep feature learning for robust representation extraction under complex conditions19. Although these advanced deep learning frameworks demonstrate strong representation learning capability, they often involve increased architectural complexity and computational cost. Motivated by these developments, the present study focuses on a Conv1D-based speech emotion recognition framework using discriminative acoustic features to achieve robust classification performance while maintaining architectural simplicity and practical applicability. One of the primary challenges in SER is the limited availability of large and balanced emotional speech datasets. Emotional speech databases often contain imbalanced distributions of emotional classes and limited speaker diversity, which can negatively affect model generalization. Additionally, many deep learning models require large datasets and complex architectures, leading to increased computational costs and making them less suitable for real-time or resource-constrained environments. Another challenge lies in the effective representation of emotional speech signals, as emotional cues are often subtle and distributed across multiple acoustic dimensions such as spectral structure, temporal dynamics, and energy variations. Consequently, achieving high recognition accuracy while maintaining computational efficiency remains a significant research challenge in speech emotion recognition. To address these challenges, this study proposes a and robust speech emotion recognition framework based on a one-dimensional convolutional neural network (1D-CNN). The proposed approach integrates discriminative acoustic features including Mel-Frequency Cepstral Coefficients (MFCC), Zero Crossing Rate (ZCR), and Root Mean Square Energy (RMSE) to capture complementary spectral and temporal characteristics of speech signals. In addition, multiple emotional speech datasets—RAVDESS, SAVEE, TESS, and CREMA-D—are combined to improve speaker diversity and enhance model generalization. To further address dataset imbalance and variability, advanced data augmentation techniques such as noise injection, time shifting, pitch modification, and time stretching are applied to expand the training dataset and improve model robustness. The architecture of the proposed 1D-CNN framework enables efficient learning of emotional speech patterns while maintaining relatively low computational complexity compared with deeper neural architectures. Despite significant progress in speech emotion recognition, existing SER frameworks still face several important challenges, including limited generalization across diverse speakers and recording conditions, dataset imbalance, high computational complexity of advanced transformer-based architectures, and insufficient robustness under real-world speech variability. Furthermore, many existing approaches rely either on highly complex deep learning architectures or on isolated acoustic features that may not effectively capture complementary emotional characteristics present in speech signals. These limitations motivate the development of a robust and efficient SER framework capable of learning meaningful temporal emotional patterns while maintaining practical applicability across multiple datasets and emotional conditions. Therefore, this study proposes a Conv1D-based speech emotion recognition framework integrating MFCC, ZCR, and RMSE acoustic features together with data augmentation and multi-dataset training to improve emotional classification performance and model generalization. Main contributions of this study The main contributions of this work are summarized as follows: A Conv1D-based speech emotion recognition framework is proposed using complementary acoustic features including MFCC, ZCR, and RMSE to capture spectral, temporal, and energy-related emotional characteristics from speech signals.A sequential feature representation strategy is introduced by concatenating frame-level acoustic descriptors into a unified one-dimensional feature vector suitable for temporal Conv1D-based learning.A combined multi-dataset SER framework utilizing RAVDESS, SAVEE, TESS, and CREMA-D datasets is developed to improve robustness and generalization capability across diverse speakers and recording conditions.Multiple audio augmentation techniques including noise injection, time shifting, pitch modification, and time stretching are integrated to improve dataset diversity and reduce class imbalance effects.Comprehensive experimental evaluation demonstrates strong classification performance across seven emotional categories with a test accuracy of 94.91% and a Macro F1-score of 0.94. A Conv1D-based speech emotion recognition framework is proposed using complementary acoustic features including MFCC, ZCR, and RMSE to capture spectral, temporal, and energy-related emotional characteristics from speech signals. A Conv1D-based speech emotion recognition framework is proposed using complementary acoustic features including MFCC, ZCR, and RMSE to capture spectral, temporal, and energy-related emotional characteristics from speech signals. A sequential feature representation strategy is introduced by concatenating frame-level acoustic descriptors into a unified one-dimensional feature vector suitable for temporal Conv1D-based learning. A sequential feature representation strategy is introduced by concatenating frame-level acoustic descriptors into a unified one-dimensional feature vector suitable for temporal Conv1D-based learning. A combined multi-dataset SER framework utilizing RAVDESS, SAVEE, TESS, and CREMA-D datasets is developed to improve robustness and generalization capability across diverse speakers and recording conditions. A combined multi-dataset SER framework utilizing RAVDESS, SAVEE, TESS, and CREMA-D datasets is developed to improve robustness and generalization capability across diverse speakers and recording conditions. Multiple audio augmentation techniques including noise injection, time shifting, pitch modification, and time stretching are integrated to improve dataset diversity and reduce class imbalance effects. Multiple audio augmentation techniques including noise injection, time shifting, pitch modification, and time stretching are integrated to improve dataset diversity and reduce class imbalance effects. Comprehensive experimental evaluation demonstrates strong classification performance across seven emotional categories with a test accuracy of 94.91% and a Macro F1-score of 0.94. Comprehensive experimental evaluation demonstrates strong classification performance across seven emotional categories with a test accuracy of 94.91% and a Macro F1-score of 0.94. Organization of the paper The remainder of this paper is organized as follows. Section 2 reviews the related literature on speech emotion recognition methods. Section 3 describes the proposed methodology, including dataset preparation, preprocessing, feature extraction, and model architecture. Section 4 presents the experimental results and performance evaluation of the proposed framework. Section 5 discusses the findings and limitations of the study, and Sect. 6 concludes the paper and outlines directions for future research. Literature review Speech emotion recognition (SER) has been widely studied as an important component of affective computing, aiming to automatically identify emotional states from speech signals. Existing approaches can broadly be categorized into three major directions: feature-based methods using handcrafted acoustic features and classical machine learning algorithms, end-to-end deep learning architectures that learn representations directly from speech signals, and methods that address data scarcity and class imbalance through augmentation strategies. Feature-based SER using handcrafted acoustic features and classical ML Early research in SER primarily relied on handcrafted acoustic features combined with traditional machine learning models. Commonly used features include Mel Frequency Cepstral Coefficients (MFCC), zero-crossing rate (ZCR), root mean square energy (RMSE), Mel spectrograms, and chroma features. These features capture spectral and temporal characteristics of speech that correlate with emotional expression. For example, K. Vamsi Krishna et al. explored SER using classifiers such as Support Vector Machines (SVM) and Multi-Layer Perceptrons (MLP) based on MFCC, MEL, and Chroma features. Their work demonstrated that discriminative acoustic features can effectively capture emotional patterns in speech signals, particularly when combined with machine learning classifiers20. Similarly, several studies emphasize the importance of sufficiently large and diverse datasets for improving the robustness and generalization ability of SER models, highlighting that dataset diversity significantly influences classification performance9. Human emotional expression in speech is inherently complex and involves multiple acoustic cues that reflect different emotional states, which motivates the continued exploration of more sophisticated feature extraction and modeling strategies17. End-to-end deep learning models (CNN/Conv1D, hybrid networks, transformers) With the advancement of deep learning, end-to-end neural architectures have become increasingly dominant in SER research because of their ability to automatically learn hierarchical feature representations from raw or minimally processed speech signals. Convolutional Neural Networks (CNNs) have been widely adopted for extracting discriminative spectral patterns from speech spectrograms. For instance, Badshah et al. applied CNNs on spectrogram representations derived from the Berlin Emotional Speech dataset and reported strong recognition performance across most emotional classes. They further demonstrated that fine-tuning a pretrained AlexNet significantly improves generalization through transfer learning21. In another study, Meng et al. proposed an advanced architecture integrating dilated CNN layers, residual connections, BiLSTM networks, and an attention mechanism to capture complex spectral–temporal dependencies in speech. Evaluated on the IEMOCAP and Berlin EMODB datasets, the ADRNN model achieved an accuracy of 63.84%, demonstrating improved feature representation capabilities22–24. Several studies have further explored hybrid deep learning architectures to capture both local and long-term dependencies in speech signals. For example, Zhao et al. proposed a hybrid CNN–LSTM architecture where convolutional layers perform feature extraction and recurrent layers model temporal relationships, resulting in improved emotion classification accuracy on benchmark datasets25. Similarly, Pepino et al. demonstrated the effectiveness of transformer-based representations using Wav2Vec2 models, which learn contextual speech representations through self-attention mechanisms and achieve strong generalization performance for SER tasks26. Attention-based transformer networks have also been investigated for identifying emotionally salient speech segments, improving classification robustness and contextual modeling27. In a related direction, Kilimci et al. showed that CNN architectures trained directly on raw audio signals can outperform traditional feature-based approaches, achieving a remarkable recognition accuracy of 99.46%28. Deep learning frameworks incorporating attention mechanisms and ensemble learning strategies have also been proposed to enhance SER performance. Lieskovsk et al. reviewed recent deep learning advances in SER and emphasized that attention mechanisms significantly improve recognition accuracy by focusing on emotionally relevant temporal and spectral cues in speech signals. However, their study also highlighted persistent challenges such as data scarcity, class imbalance, and cross-corpus generalization29. Similarly, Ahmed et al. proposed a weighted ensemble deep learning framework incorporating both local and global attention blocks, demonstrating improved robustness and generalization across multiple languages and datasets when combined with data augmentation techniques30. Data augmentation and class-imbalance handling Another significant research direction focuses on data augmentation and strategies for addressing dataset imbalance, which is a common issue in SER datasets due to uneven distribution of emotional categories. Majd Saloumi et al. developed a 1D-CNN based SER model using MFCC features for short speech recordings and demonstrated that applying data augmentation techniques significantly improves recognition accuracy, achieving 83% accuracy on the RAVDESS dataset31. Likewise, Baek and Lee proposed a DCGAN-based data augmentation approach that generates synthetic emotional speech samples to address data scarcity and class imbalance, leading to improved model robustness compared with conventional CNN-based SER models4. Beyond speech signals, research in affective computing has also explored multimodal and cross-domain emotion recognition approaches. Deep convolutional neural networks have been successfully applied to sensor-based activity recognition tasks, illustrating the ability of deep learning models to extract discriminative patterns from complex temporal data streams32. Similarly, swarm-intelligence-based feature and channel selection techniques have been used for EEG-based emotion recognition, highlighting the importance of feature optimization in affective computing systems33. Other studies have proposed advanced architectures such as deep sparse autoencoder–extreme learning machine models and quaternion nonlinear echo state networks to enhance spectro-temporal feature representation and improve classification performance in SER tasks34,35. Emotion and mental health detection have also been investigated using EEG-based neural frameworks. Machine learning models have been developed for detecting major depressive disorder and bipolar disorder from EEG signals, demonstrating the potential of artificial intelligence in clinical decision support systems36. More recent work has explored multimodal approaches combining audiovisual and EEG data to improve depression recognition and relapse prediction37. Additionally, investigations into both active and passive EEG paradigms have provided deeper insights into neural activity patterns associated with depression and other mental health conditions38. Recent studies have further expanded SER research toward multimodal emotion recognition and transformer-based architectures. The TEMPO framework introduces training-time modality balancing to improve multimodal sentiment analysis by mitigating modality dominance during optimization39. Diffusion-transformer architectures such as RMER-DT have also been proposed to model cross-modal dependencies in conversational emotion recognition tasks40. Other approaches integrate audio, visual, and textual modalities using contrastive learning strategies to enhance multimodal emotion analysis41. Transformer-based adversarial fusion strategies such as RAFT further improve multimodal sentiment understanding through robust feature integration42. Similarly, context-aware frameworks such as CIME leverage contextual interactions between modalities to improve semantic fusion for emotion recognition43. Advanced neural architectures incorporating deformable non-local attention and multi-branch feature fusion have also been successfully applied to visual tasks such as image super-resolution, demonstrating the versatility of deep learning frameworks in complex pattern recognition problems44. More recently, hybrid architectures combining convolutional neural networks with transformer-based attention mechanisms have shown promising results. For example, Das et al. proposed a hybrid SER model integrating 1D-CNNs with a multi-head attention transformer using acoustic features such as ZCR, MFCC, and RMSE. Their approach demonstrated strong cross-lingual performance across multiple benchmark datasets, achieving up to 96% accuracy and outperforming several existing methods17. Similarly, Dasude et al. investigated deep learning architectures including CNNs, LSTMs, and ConvLSTMs, showing that combining spatial and temporal feature extraction mechanisms significantly improves emotion recognition performance, achieving 50.6% accuracy on a combined dataset consisting of TESS, RAVDESS, SAVEE, and CREMA-D corpora45. Overall, these studies highlight the rapid evolution of SER research from handcrafted feature-based approaches to advanced deep learning and multimodal architectures. Despite these advances, many existing methods rely on complex models or multimodal data sources that require significant computational resources. In contrast, the present study focuses on a computationally efficient speech-based emotion recognition framework that integrates discriminative acoustic features with a deep learning architecture, aiming to achieve robust recognition performance while maintaining practical applicability in real-world systems46,47. Table 1 summarizes the key studies related to speech emotion recognition (SER), highlighting the datasets used, feature extraction techniques, model architectures, and reported performance. In addition, the limitations of each approach are identified to provide a clearer understanding of the existing research gaps and motivate the proposed framework. Table 1Comparative summary of representative speech emotion recognition studies.ReferencesDatasetFeature ExtractionModelResultsWeaknesses 21 Berlin EMO-DBSpectrogram featuresCNN with transfer learning (AlexNet)High recognition accuracy for most emotionsLower performance for certain emotions such as fear; limited dataset size 22–24 IEMOCAP, EMO-DBSpectral–temporal featuresADRNN (Dilated CNN + Residual Blocks + BiLSTM + Attention)Achieved 63.84% accuracyComplex architecture with high computational cost 24 Multiple SER datasetsAcoustic and spectral featuresAttention-based deep learning modelsImproved recognition using attention mechanismsData scarcity and cross-corpus generalization issues 29 Multilingual SER datasetsSpectrogram-based featuresWeighted ensemble deep learning with attentionState-of-the-art performance with strong generalizationHigh training complexity and resource requirements 28 SER benchmark datasetsRaw speech signalsEnd-to-end CNNAchieved up to 99.46% accuracyPotential overfitting and dependency on dataset characteristics 31 RAVDESSMFCC1D-CNNAchieved 83% accuracy with augmentationLimited dataset size; performance sensitive to data diversity 20 SER speech datasetsMFCC, MEL, ChromaSVM, MLP, CNNImproved classification accuracy using hybrid featuresRequires manual feature engineering 4 SER datasetsSpectrogram featuresDCGAN + CNNImproved robustness through data augmentationGAN training instability and synthetic data reliability concerns 25 SER benchmark datasetsSpectrogram featuresCNN–LSTM hybridImproved modeling of temporal dependenciesIncreased model complexity 26 SER datasetsRaw speech representationsWav2Vec2 transformer modelStrong generalization performanceRequires large pretrained models and high computational cost 27 SER datasetsSpectral featuresAttention-based transformer networkImproved emotion classification robustnessLarge data requirements 17 Multiple SER datasetsZCR, MFCC, RMSE1D-CNN + Multi-Head Attention TransformerAchieved 96% accuracy with cross-lingual performanceHigher computational complexity 45 TESS, RAVDESS, SAVEE, CREMA-DSpectrogram featuresCNN, LSTM, ConvLSTMAchieved 50.6% accuracy on combined datasetModerate accuracy due to dataset heterogeneity Table 1Comparative summary of representative speech emotion recognition studies.ReferencesDatasetFeature ExtractionModelResultsWeaknesses 21 Berlin EMO-DBSpectrogram featuresCNN with transfer learning (AlexNet)High recognition accuracy for most emotionsLower performance for certain emotions such as fear; limited dataset size 22–24 IEMOCAP, EMO-DBSpectral–temporal featuresADRNN (Dilated CNN + Residual Blocks + BiLSTM + Attention)Achieved 63.84% accuracyComplex architecture with high computational cost 24 Multiple SER datasetsAcoustic and spectral featuresAttention-based deep learning modelsImproved recognition using attention mechanismsData scarcity and cross-corpus generalization issues 29 Multilingual SER datasetsSpectrogram-based featuresWeighted ensemble deep learning with attentionState-of-the-art performance with strong generalizationHigh training complexity and resource requirements 28 SER benchmark datasetsRaw speech signalsEnd-to-end CNNAchieved up to 99.46% accuracyPotential overfitting and dependency on dataset characteristics 31 RAVDESSMFCC1D-CNNAchieved 83% accuracy with augmentationLimited dataset size; performance sensitive to data diversity 20 SER speech datasetsMFCC, MEL, ChromaSVM, MLP, CNNImproved classification accuracy using hybrid featuresRequires manual feature engineering 4 SER datasetsSpectrogram featuresDCGAN + CNNImproved robustness through data augmentationGAN training instability and synthetic data reliability concerns 25 SER benchmark datasetsSpectrogram featuresCNN–LSTM hybridImproved modeling of temporal dependenciesIncreased model complexity 26 SER datasetsRaw speech representationsWav2Vec2 transformer modelStrong generalization performanceRequires large pretrained models and high computational cost 27 SER datasetsSpectral featuresAttention-based transformer networkImproved emotion classification robustnessLarge data requirements 17 Multiple SER datasetsZCR, MFCC, RMSE1D-CNN + Multi-Head Attention TransformerAchieved 96% accuracy with cross-lingual performanceHigher computational complexity 45 TESS, RAVDESS, SAVEE, CREMA-DSpectrogram featuresCNN, LSTM, ConvLSTMAchieved 50.6% accuracy on combined datasetModerate accuracy due to dataset heterogeneity Comparative summary of representative speech emotion recognition studies. Although recent speech emotion recognition studies have demonstrated improved performance using CNNs, recurrent neural networks, attention mechanisms, and transformer-based architectures, several important research gaps still remain. Many existing SER frameworks rely on computationally complex architectures requiring extensive training data and high computational resources, which may limit their practical applicability in real-world environments. In addition, several studies focus primarily on single-dataset evaluation, resulting in limited generalization capability across diverse speakers, recording conditions, and emotional variations. Existing approaches also frequently utilize isolated acoustic features or highly complex end-to-end representations without adequately exploring complementary feature integration strategies. Furthermore, challenges related to dataset imbalance, robustness under noisy conditions, and efficient temporal feature learning continue to affect SER performance. Motivated by these limitations, the present study proposes a Conv1D-based speech emotion recognition framework integrating MFCC, ZCR, and RMSE acoustic features together with data augmentation and multi-dataset evaluation to improve emotional classification robustness and generalization capability. Methodology Understanding the fundamentals of speech emotion recognition (SER) systems is crucial to appreciating the advancements made in this study. In order for readers to comprehend the prerequisites required for the methods used in this paper, this section provides an overview of the main ideas, characteristics, and approaches related to proposed model. Workflow Speech Emotion Recognition Process: There are two primary steps are involved in the SER process. Feature extraction To extract important features that convey emotional information, speech signals are pre-processed. Mel-frequency and other acoustic, prosodic, and spectral parameters are commonly included in these features. Harmonic to Noise Ratio (HNR), Zero Crossing Rate (ZCR), Cepstral Coefficients (MFCC), and Teager Energy Operator (TEO). Emotion classification Machine learning or deep learning models categorize the speech signal into predetermined emotion categories based on the extracted features. By utilizing unlabeled speech data, multitask self-supervised learning improves speech emotion recognition and increases robust-ness across a range of speakers, emotions, and environmental conditions25,30. Fig. 1Workflow of the proposed speech emotion recognition (SER) framework. Workflow of the proposed speech emotion recognition (SER) framework. Figure 1 The proposed framework presents a structured pipeline for speech emotion recognition. To ensure data integrity, scale verification, and consistency across samples, data preprocessing is first carried out. In order to reduce bias in the classification of emotions, class and data imbalance are then evaluated. Time-frequency characteristics of speech signals are captured through multimodal feature exploration using spectrogram and wavelet representations. Data augmentation techniques, such as noise addition, time shifting, pitch modification, and time stretching, improve the robustness of the dataset. To accurately depict emotional cues, acoustic characteristics like Root Mean Square Energy, Mel-Frequency Cepstral Coefficients, and Zero-Crossing Rate are extracted. Feature analysis, data splitting, standardization, and reshaping for effective learning are all part of the model-building process. Accuracy, precision, recall, and F1-score are used to assess the model’s performance. The proposed framework receives raw speech audio signals from the combined SER datasets as input. The audio signals first undergo preprocessing steps including normalization, resampling, padding/truncation, and augmentation for training samples. Subsequently, discriminative acoustic features including MFCC, ZCR, and RMSE are extracted to form a unified feature representation. These feature vectors are then provided as input to the Conv1D-based deep learning architecture for hierarchical temporal feature learning. Finally, the Softmax output layer predicts one of the seven target emotional categories: disgust, happy, sad, neutral, fear, angry, and surprise. Dataset Although it directly affects both training performance and generalization capacity, the choice of a suitable dataset is crucial to the success of speech emotion recognition (SER) systems. To guarantee robustness in practical applications, an ideal dataset should reflect a broad variety of emotional expressions, demographic diversity, and speaking styles21. Nevertheless, there are still few publicly accessible datasets with this level of thorough variety. This study uses a number of popular SER datasets, such as RAVDESS, TESS, SAVEE, and CREMA-D, to assess the suggested method to overcome this difficulty. The suggested approach uses spectrogram and mel-spectrogram representations taken from brief voice recordings to identify emotions22,30. Convolutional neural networks (CNNs) mainly use spectrograms, mel-spectrograms, or raw audio signals as input characteristics, according to earlier research22. To ascertain whether feature representation produces better performance, spectrograms and mel-spectrograms are jointly studied in this work. This study makes use of several datasets and their combinations, in contrast to many other works that rely on a single dataset48,49. This allows for a more thorough evaluation and improved understanding of the generalization capabilities of SER systems in actual circumstances. RAVDESS: A popular dataset in studies on voice emotion identification is the (Ryerson Audio-Visual Database of Emotional voice and Song)50. Eight distinct speech expressions and five distinct song expressions for the two sentences in the North American dialect are recorded by 24 professional performers (12 male and 12 female). One of eight emotions—neutral, peaceful, happy, sad, angry, afraid, astonished, and disgusted—is assigned to each audio sample, which lasts roughly three seconds. These emotions are expressed at two strength levels (normal and strong). Only the speech recordings were used in this investigation. Despite its widespread use, RAVDESS’s smaller size makes it inadequate for deep neural network training on its own.TESS: TESS (The Toronto Emotional Speech Set)51 is a female-only dataset, two female speakers, ages 26 and 64, provided 2800 audio phrases for the English speech emotion recognition dataset known as TESS. To ensure phonetic consistency, 200 target words are uttered within a pre-determined carrier phrase in the recordings. Seven categories of emotions are represented in the dataset: fear, anger, disgust, happiness, neutrality, sadness, and surprise.SAVEE: The SAVEE Dataset (Surrey Audio–Visual Expressed Emotion) is a high-quality audio compilation of speech from men only. There are 480 audio files in the SAVEE file. Sixty words are employed for more emotions and 120 words for negative emotions. Seven emotions can be found in this document: joy, fear, sadness, neutrality, amazement, disgust, and rage. This could be in line with another data set that is exclusive to men52.CREMA-D: The crowd-sourced emotional multimodal actors dataset is known as CREMA-D53. It displays multimodal data, which is data that corresponds to both visual and auditory modalities. 91 actors participated in the preparation of the 7443 clips that make up the dataset. This dataset is trustworthy since 2443 raters provided categorical emotion labels through crowdsourcing. Neutral, joy, anger, contempt, fear, and sadness are the emotions depicted in CREMA-D. Neural networks can be trained just on this dataset thanks to such emotional examples from several actors. RAVDESS: A popular dataset in studies on voice emotion identification is the (Ryerson Audio-Visual Database of Emotional voice and Song)50. Eight distinct speech expressions and five distinct song expressions for the two sentences in the North American dialect are recorded by 24 professional performers (12 male and 12 female). One of eight emotions—neutral, peaceful, happy, sad, angry, afraid, astonished, and disgusted—is assigned to each audio sample, which lasts roughly three seconds. These emotions are expressed at two strength levels (normal and strong). Only the speech recordings were used in this investigation. Despite its widespread use, RAVDESS’s smaller size makes it inadequate for deep neural network training on its own. RAVDESS: A popular dataset in studies on voice emotion identification is the (Ryerson Audio-Visual Database of Emotional voice and Song)50. Eight distinct speech expressions and five distinct song expressions for the two sentences in the North American dialect are recorded by 24 professional performers (12 male and 12 female). One of eight emotions—neutral, peaceful, happy, sad, angry, afraid, astonished, and disgusted—is assigned to each audio sample, which lasts roughly three seconds. These emotions are expressed at two strength levels (normal and strong). Only the speech recordings were used in this investigation. Despite its widespread use, RAVDESS’s smaller size makes it inadequate for deep neural network training on its own. TESS: TESS (The Toronto Emotional Speech Set)51 is a female-only dataset, two female speakers, ages 26 and 64, provided 2800 audio phrases for the English speech emotion recognition dataset known as TESS. To ensure phonetic consistency, 200 target words are uttered within a pre-determined carrier phrase in the recordings. Seven categories of emotions are represented in the dataset: fear, anger, disgust, happiness, neutrality, sadness, and surprise. TESS: TESS (The Toronto Emotional Speech Set)51 is a female-only dataset, two female speakers, ages 26 and 64, provided 2800 audio phrases for the English speech emotion recognition dataset known as TESS. To ensure phonetic consistency, 200 target words are uttered within a pre-determined carrier phrase in the recordings. Seven categories of emotions are represented in the dataset: fear, anger, disgust, happiness, neutrality, sadness, and surprise. SAVEE: The SAVEE Dataset (Surrey Audio–Visual Expressed Emotion) is a high-quality audio compilation of speech from men only. There are 480 audio files in the SAVEE file. Sixty words are employed for more emotions and 120 words for negative emotions. Seven emotions can be found in this document: joy, fear, sadness, neutrality, amazement, disgust, and rage. This could be in line with another data set that is exclusive to men52. SAVEE: The SAVEE Dataset (Surrey Audio–Visual Expressed Emotion) is a high-quality audio compilation of speech from men only. There are 480 audio files in the SAVEE file. Sixty words are employed for more emotions and 120 words for negative emotions. Seven emotions can be found in this document: joy, fear, sadness, neutrality, amazement, disgust, and rage. This could be in line with another data set that is exclusive to men52. CREMA-D: The crowd-sourced emotional multimodal actors dataset is known as CREMA-D53. It displays multimodal data, which is data that corresponds to both visual and auditory modalities. 91 actors participated in the preparation of the 7443 clips that make up the dataset. This dataset is trustworthy since 2443 raters provided categorical emotion labels through crowdsourcing. Neutral, joy, anger, contempt, fear, and sadness are the emotions depicted in CREMA-D. Neural networks can be trained just on this dataset thanks to such emotional examples from several actors. CREMA-D: The crowd-sourced emotional multimodal actors dataset is known as CREMA-D53. It displays multimodal data, which is data that corresponds to both visual and auditory modalities. 91 actors participated in the preparation of the 7443 clips that make up the dataset. This dataset is trustworthy since 2443 raters provided categorical emotion labels through crowdsourcing. Neutral, joy, anger, contempt, fear, and sadness are the emotions depicted in CREMA-D. Neural networks can be trained just on this dataset thanks to such emotional examples from several actors. Table 2 presents the distribution of emotional speech samples across the four datasets used in this study, including RAVDESS, TESS, SAVEE, and CREMA-D. The table summarizes the number of samples for each emotion category, providing an overview of the dataset composition used for training and evaluation of the proposed SER model. Table 2Emotion-wise sample distribution of SER across different datasets.EmotionTESSRAVDESSSAVEECREMA-DDisgust40019260893Happy400192601265Sad400192601116Neutral4002881201637Fear400192601190Angry400192601339Surprise400192600Overall280014404807442 Table 2Emotion-wise sample distribution of SER across different datasets.EmotionTESSRAVDESSSAVEECREMA-DDisgust40019260893Happy400192601265Sad400192601116Neutral4002881201637Fear400192601190Angry400192601339Surprise400192600Overall280014404807442 Emotion-wise sample distribution of SER across different datasets. To improve the diversity of the training data, the four emotional speech datasets—RAVDESS, TESS, SAVEE, and CREMA-D—were combined into a unified dataset before model training. The original datasets contain 2,800 samples from TESS, 1,440 from RAVDESS, 480 from SAVEE, and 7,442 from CREMA-D. Since the emotional categories across these datasets differ slightly, the samples were mapped into a common set of seven emotion classes. After preprocessing and class alignment, a subset of the available recordings was used to maintain class balance across the merged dataset. The combined dataset was then expanded using four audio augmentation techniques—noise injection, time shifting, pitch modification, and time stretching—to increase variability in speech signals. The combined dataset obtained from RAVDESS, TESS, SAVEE, and CREMA-D consisted of 12,162 samples. The dataset was divided into training, validation, and testing subsets using a 72:8:20 ratio, resulting in 8,756 training samples, 973 validation samples, and 2,433 testing samples. Data augmentation techniques were applied only to the training subset, increasing the effective training samples to 11,682. Feature extraction using MFCC, ZCR, and RMSE was then performed on the processed audio samples. Fig. 2Sample distribution across emotion categories after dataset fusion and the corresponding counts of the seven concatenated emotion classes in the SER dataset. Sample distribution across emotion categories after dataset fusion and the corresponding counts of the seven concatenated emotion classes in the SER dataset. Pre-processing Acoustic characterization One sample of each emotion is subjected to the wave plot and spectrogram algorithms. This stage verifies the extractability of features by offering tangible visual proof of the acoustic distinctions between emotions (such as the high-frequency energy of ”fear” against the low-frequency energy of ”sadness”). The Figs. 3, 4 and 5 analysis reveals distinct acoustic differences between the emotions of disgust and fear. Speech associated with disgust is marked by a dominance of low frequencies, irregular spectral patterns, and controlled yet tense amplitude variations, indicating a restricted vocal effort. In contrast, speech related to fear shows increased and unstable energy in the low-to-mid frequency range, suggesting heightened arousal. Rapid spectral changes and sudden temporal shifts further highlight vocal instability in fear. Wave-form analysis indicates moderate tension in disgust compared to large, uneven amplitude peaks in fear. Overall, disgust is characterized by controlled tension, while fear is associated with extreme arousal and prosodic instability. Fig. 3Spectrogram and waveform representation of the disgust emotion. Spectrogram and waveform representation of the disgust emotion. Fig. 4Spectrogram and waveform representation of the angry emotion. Spectrogram and waveform representation of the angry emotion. The comparative study in Figs. 4 and 6 reveals distinct acoustic distinctions between angry and happy emotions. Angry speech is marked by concentrated spec-tral energy, quick temporal changes, and significant waveform amplitudes, indicating high arousal and vocal strain. On the other hand, happy speech features widely spread spectral energy with more consistent patterns, suggest-ing expressive yet controlled vocal delivery. The waveform variations in happy speech are rhythmic and moderate, contrasting with the intense peaks found in angry speech. In essence, anger communicates dominance and intensity, while happiness conveys positive arousal and vocal expressiveness. Fig. 5Spectrogram and waveform representation of the fear emotion. Spectrogram and waveform representation of the fear emotion. Fig. 6Spectrogram and waveform representation of the happy emotion. Spectrogram and waveform representation of the happy emotion. The illustration 7 and 8 contrasts the acoustic features of surprise and neutral emotions through spectrogram and waveform depictions. Surprise speech is characterized by sudden bursts of spectral energy and pronounced waveform peaks, indicating rapid vocal excitement and brief utterances. On the other hand, neutral speech is marked by steady, evenly distributed spectral energy with smooth temporal changes. The neutral waveform displays moderate and consistent amplitude without sudden spikes, suggesting controlled loudness. In summary, surprise emotion is distinguished by high characterized, while neutral emotion is characterized by acoustic stability and minimal prosodic modulation. The Fig. 9 depicts the acoustic properties of sadness through spectrogram and waveform visuals. The spectrogram reveals low-intensity energy concentrated in the lower frequency ranges, suggesting a decrease in vocal effort. The limited spectral variation indicates slow and subdued speech dynamics. The waveform displays low amplitude and smooth temporal changes, which align with soft and monotonous speech. Collectively, these features highlight the low-arousal and restrained prosodic characteristics of sad speech. Fig. 7Spectrogram and waveform representation of the neutral emotion. Spectrogram and waveform representation of the neutral emotion. Fig. 8Spectrogram and waveform representation of the surprise emotion. Spectrogram and waveform representation of the surprise emotion. Fig. 9Spectrogram and waveform representation of the sad emotion. Spectrogram and waveform representation of the sad emotion. Data augmentation and optimization In situations where training data is insufficient, data augmentation techniques are essential for enhancing the resilience and generalizability of machine learning models. Data augmentation modifies audio signals to produce a range of training examples while preserving their underlying emotional content in the context of speech emotion recognition28. Figure 10 is the original sample of voice is extracted from the datasets. Noise Injection: In the Fig. 11 to increase the model’s resistance to noise interference and replicate real-world environmental fluctuations, random noise is added to the audio output.Shifting: Vocal pitch fluctuations are produced by altering the audio signal’s pitch without altering the emotional content. This method mimics shifts in speaker identities or vocal characteristics. Figure 12 illustrates the time-shifting augmentation applied to the speech signal, which simulates temporal variations in the audio sample.Pitching: Vocal pitch fluctuations are produced by altering the audio signal’s pitch without altering the emotional content. This method mimics shifts in speaker identities or vocal characteristics. Figure 13 shows the pitch modification applied to the speech signal.Stretching: Changes an audio signal’s length without changing its pitch or emotional qualities. This method mimics varying speech rates or emotional expression durations by using temporal fluctuations. The Fig. 14 simulates variations in speaking rate, ensuring the model remains accurate even if a person speaks quickly (compressed time) or slowly (stretched time) while expressing the same emotion. Noise Injection: In the Fig. 11 to increase the model’s resistance to noise interference and replicate real-world environmental fluctuations, random noise is added to the audio output. Noise Injection: In the Fig. 11 to increase the model’s resistance to noise interference and replicate real-world environmental fluctuations, random noise is added to the audio output. Shifting: Vocal pitch fluctuations are produced by altering the audio signal’s pitch without altering the emotional content. This method mimics shifts in speaker identities or vocal characteristics. Figure 12 illustrates the time-shifting augmentation applied to the speech signal, which simulates temporal variations in the audio sample. Shifting: Vocal pitch fluctuations are produced by altering the audio signal’s pitch without altering the emotional content. This method mimics shifts in speaker identities or vocal characteristics. Figure 12 illustrates the time-shifting augmentation applied to the speech signal, which simulates temporal variations in the audio sample. Pitching: Vocal pitch fluctuations are produced by altering the audio signal’s pitch without altering the emotional content. This method mimics shifts in speaker identities or vocal characteristics. Figure 13 shows the pitch modification applied to the speech signal. Pitching: Vocal pitch fluctuations are produced by altering the audio signal’s pitch without altering the emotional content. This method mimics shifts in speaker identities or vocal characteristics. Figure 13 shows the pitch modification applied to the speech signal. Stretching: Changes an audio signal’s length without changing its pitch or emotional qualities. This method mimics varying speech rates or emotional expression durations by using temporal fluctuations. The Fig. 14 simulates variations in speaking rate, ensuring the model remains accurate even if a person speaks quickly (compressed time) or slowly (stretched time) while expressing the same emotion. Stretching: Changes an audio signal’s length without changing its pitch or emotional qualities. This method mimics varying speech rates or emotional expression durations by using temporal fluctuations. The Fig. 14 simulates variations in speaking rate, ensuring the model remains accurate even if a person speaks quickly (compressed time) or slowly (stretched time) while expressing the same emotion. Fig. 10Example waveform of an original speech signal before augmentation. Example waveform of an original speech signal before augmentation. Fig. 11Speech waveform after applying noise injection augmentation. Speech waveform after applying noise injection augmentation. Fig. 12Speech waveform after applying time-shifting augmentation. Speech waveform after applying time-shifting augmentation. Fig. 13Speech waveform after applying pitch modification augmentation. Speech waveform after applying pitch modification augmentation. Fig. 14Speech waveform after applying time-stretching augmentation. Speech waveform after applying time-stretching augmentation. To increase dataset diversity and improve the robustness of the model, several audio data augmentation techniques were applied to the training samples. Noise injection was performed by adding low-level Gaussian noise to the audio signal with a noise factor in the range of 0.001–0.005 of the signal amplitude, simulating environmental background noise. Time shifting was implemented by randomly shifting the waveform along the time axis by ± 0.2 s, allowing the model to learn emotion patterns independent of exact temporal alignment. Pitch modification was applied by altering the pitch of the audio signal within a range of ± 2 semitones, simulating variations in speaker vocal characteristics while preserving emotional content. Time stretching was performed with a stretching factor between 0.8 and 1.2, which simulates variations in speaking speed without changing pitch. These augmentation techniques were applied randomly during preprocessing to expand the effective training dataset and improve model generalization. These augmentation techniques were applied to samples from all four datasets to increase the diversity of emotional speech signals and contribute to the final combined dataset used for training the proposed model. To prevent data leakage and ensure fair model evaluation, data augmentation was applied only to the training dataset after dataset partitioning. The validation and testing datasets remained completely independent and were not augmented. Feature extraction and selection Feature extraction plays a crucial role in speech emotion recognition (SER) because emotional information in speech signals is often embedded in subtle acoustic variations such as pitch, energy, and spectral characteristics. Effective feature extraction techniques help transform raw audio signals into representative numerical features that can capture the emotional characteristics of speech. In this study, three widely used acoustic features—Mel-Frequency Cepstral Coefficients (MFCC), Zero-Crossing Rate (ZCR), and Root Mean Square Energy (RMSE)—are extracted from speech signals to represent complementary spectral, temporal, and energy-related properties of emotional speech. MFCCs: Mel-frequency cepstral coefficients (MFCC) are among the most commonly used features in speech processing and emotion recognition because they effectively represent the spectral characteristics of speech signals in a way that closely resembles human auditory perception. The MFCC extraction process involves several steps, including pre-emphasis, framing, windowing, Fourier transformation, Mel-scale filter bank processing, and discrete cosine transformation. The Mel scale converts the frequency axis into a perceptual scale that reflects how humans perceive pitch differences. By applying a bank of Mel-spaced filters to the power spectrum of the speech signal, MFCC features capture the spectral envelope of the speech signal, which is strongly related to vocal tract configuration and emotional expression. MFCCs are the most important attributes for speech processing54. MFCCs: Mel-frequency cepstral coefficients (MFCC) are among the most commonly used features in speech processing and emotion recognition because they effectively represent the spectral characteristics of speech signals in a way that closely resembles human auditory perception. The MFCC extraction process involves several steps, including pre-emphasis, framing, windowing, Fourier transformation, Mel-scale filter bank processing, and discrete cosine transformation. The Mel scale converts the frequency axis into a perceptual scale that reflects how humans perceive pitch differences. By applying a bank of Mel-spaced filters to the power spectrum of the speech signal, MFCC features capture the spectral envelope of the speech signal, which is strongly related to vocal tract configuration and emotional expression. MFCCs are the most important attributes for speech processing54. MFCCs: Mel-frequency cepstral coefficients (MFCC) are among the most commonly used features in speech processing and emotion recognition because they effectively represent the spectral characteristics of speech signals in a way that closely resembles human auditory perception. The MFCC extraction process involves several steps, including pre-emphasis, framing, windowing, Fourier transformation, Mel-scale filter bank processing, and discrete cosine transformation. The Mel scale converts the frequency axis into a perceptual scale that reflects how humans perceive pitch differences. By applying a bank of Mel-spaced filters to the power spectrum of the speech signal, MFCC features capture the spectral envelope of the speech signal, which is strongly related to vocal tract configuration and emotional expression. MFCCs are the most important attributes for speech processing54. Mathematically, the Mel-frequency scale is defined as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{*{20}c} {m\left( f \right) = 2595{\mathrm{log}}_{{10}} \left( {1 + \frac{f}{{700}}} \right)} \\ \end{array}$$\end{document} where f represents the frequency in Hertz. The resulting MFCC coefficients provide a compact representation of speech spectral properties and have been widely used for tasks such as speech recognition, speaker identification, and emotion recognition. ZCR: The Zero Crossing Rate (ZCR) is a time-domain feature that measures how frequently the audio signal waveform crosses the zero amplitude axis within a given frame. It reflects the rate at which the signal changes sign from positive to negative or vice versa. ZCR is particularly useful for capturing the temporal structure of speech signals and distinguishing between voiced and unvoiced speech components. Emotional speech often exhibits variations in articulation, intensity, and vocal tension, which can influence the zero-crossing behavior of the signal55. ZCR: The Zero Crossing Rate (ZCR) is a time-domain feature that measures how frequently the audio signal waveform crosses the zero amplitude axis within a given frame. It reflects the rate at which the signal changes sign from positive to negative or vice versa. ZCR is particularly useful for capturing the temporal structure of speech signals and distinguishing between voiced and unvoiced speech components. Emotional speech often exhibits variations in articulation, intensity, and vocal tension, which can influence the zero-crossing behavior of the signal55. ZCR: The Zero Crossing Rate (ZCR) is a time-domain feature that measures how frequently the audio signal waveform crosses the zero amplitude axis within a given frame. It reflects the rate at which the signal changes sign from positive to negative or vice versa. ZCR is particularly useful for capturing the temporal structure of speech signals and distinguishing between voiced and unvoiced speech components. Emotional speech often exhibits variations in articulation, intensity, and vocal tension, which can influence the zero-crossing behavior of the signal55. The ZCR for a signal frame is calculated as: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ZCR = \frac{1}{{N - 1}}\sum {\:_{{t = 1}}^{{N_{1} }} } {\mathbb{K}}_{{R < 0}} \left( {S_{n} \left( t \right)S_{n} \left( {t - 1} \right)} \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{n}$$\end{document} represents the signal amplitude at time n, and N is the number of samples in the frame. Higher ZCR values are typically associated with unvoiced speech or noisy signals, while lower values are characteristic of voiced speech segments. RMSE: Root mean square energy (RMSE) measures the average energy of the speech signal within a frame and provides information about the intensity or loudness of speech. Emotional states such as anger, excitement, or fear often exhibit higher energy levels, whereas emotions such as sadness or neutrality typically show lower energy levels. Therefore, energy-based features are important indicators of emotional expression in speech signals11. RMSE: Root mean square energy (RMSE) measures the average energy of the speech signal within a frame and provides information about the intensity or loudness of speech. Emotional states such as anger, excitement, or fear often exhibit higher energy levels, whereas emotions such as sadness or neutrality typically show lower energy levels. Therefore, energy-based features are important indicators of emotional expression in speech signals11. RMSE: Root mean square energy (RMSE) measures the average energy of the speech signal within a frame and provides information about the intensity or loudness of speech. Emotional states such as anger, excitement, or fear often exhibit higher energy levels, whereas emotions such as sadness or neutrality typically show lower energy levels. Therefore, energy-based features are important indicators of emotional expression in speech signals11. The RMSE value of a speech frame is computed as:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{*{20}c} {RMSE_{a} = \sqrt {\frac{1}{n}\left( {a_{1}^{2} + a_{2}^{2} + a_{3}^{2} + \ldots \: \ldots \: \ldots \:..a_{n}^{2} } \right)} } \\ \end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{n}$$\end{document} represents the amplitude of the speech signal and N denotes the number of samples in the frame. Feature combination Each of these features captures different aspects of emotional speech signals. MFCC features describe the spectral structure of speech, ZCR reflects temporal signal transitions, and RMSE represents signal energy. By combining these complementary features, the proposed model can effectively capture both spectral and temporal variations associated with different emotional states. For each 3-second audio sample, frame-level MFCC, ZCR, and RMSE features were extracted using a frame length of 25 ms and a hop length of 10 ms. The resulting frame-wise acoustic descriptors were sequentially concatenated to form a unified one-dimensional feature vector suitable for Conv1D processing. In this study, the extracted MFCC, ZCR, and RMSE features are concatenated into a unified feature representation that serves as the input to the proposed deep learning model. The final input representation consists of 40 MFCC coefficients together with 1 ZCR and 1 RMSE descriptor extracted across multiple frames, resulting in a flattened sequential input tensor of size (2376, 1). Since ZCR and RMSE individually provide only low-dimensional scalar information, standalone Conv1D-based learning using these features alone was not effective within the proposed 1D-CNN framework. In preliminary experiments, the network could not learn sufficiently discriminative representations using only ZCR or RMSE because these descriptors do not contain rich temporal-spectral information required for convolutional feature extraction. Therefore, ZCR and RMSE were utilized as complementary auxiliary features alongside MFCC representations to improve emotional discrimination and overall model robustness. This combined representation improves the discriminative capability of the model and enhances emotion classification performance. Audio preprocessing and feature extraction settings All audio recordings were resampled to a sampling rate of 22,050 Hz to ensure consistency across the datasets. Each audio clip was standardized to a fixed duration of 3 s. Signals shorter than this duration were zero-padded, while longer signals were truncated to maintain uniform input length. The audio signals were then normalized to the range [–1,1] to stabilize model training. For feature extraction, Mel-Frequency Cepstral Coefficients (MFCCs) were computed using 40 coefficients with a frame length of 25 ms and a hop length of 10 ms, using a Hamming window function. In addition to MFCCs, Zero-Crossing Rate (ZCR) and Root Mean Square Energy (RMSE) features were extracted from the same frames to capture complementary temporal and energy-related characteristics of the speech signals. To increase dataset diversity and improve model robustness, several audio data augmentation techniques were applied. Noise injection was performed by adding Gaussian noise with an amplitude factor in the range of 0.001–0.005 of the signal amplitude. Time shifting was implemented by randomly shifting the audio signal within a range of ± 0.2 s. Pitch modification was applied by adjusting the pitch within a range of ± 2 semitones, and time stretching was performed using stretch factors between 0.8 and 1.2. These augmentations simulate variations in real-world speech signals while preserving the emotional characteristics of the audio. Feature extraction was performed after dataset partitioning and augmentation of the training subset. The testing dataset remained independent throughout the training process. The extracted frame-level MFCC, ZCR, and RMSE features were reshaped into a one-dimensional sequential vector representation before being provided to the Conv1D network. The resulting input tensor size (2376, 1) corresponds to the concatenated acoustic feature sequence extracted from each audio sample. Proposed architectural model The design of the proposed model is motivated by the temporal nature of speech signals and the need for a computationally efficient architecture capable of capturing emotional dynamics in audio sequences. Unlike more complex deep learning architectures such as transformer-based models or graph neural networks, which typically require large-scale datasets and higher computational resources, the proposed one-dimensional convolutional neural network (1D-CNN) provides an effective balance between model complexity and classification performance. The use of stacked Conv1D layers enables the model to capture local temporal dependencies and short-term acoustic variations such as pitch fluctuations, energy bursts, and spectral transitions that are strongly correlated with emotional expressions in speech. Batch normalization and max-pooling layers further stabilize training and improve feature abstraction while reducing overfitting. The architecture was intentionally designed to remain so that it can be applied to real-time or resource-constrained environments without sacrificing predictive accuracy. To validate the effectiveness of the proposed architecture, comparative experiments with baseline machine learning models and previously reported CNN-based methods were conducted, demonstrating that the proposed model achieves competitive performance while maintaining computational efficiency. This design choice highlights that an appropriately structured 1D-CNN combined with discriminative acoustic features and data augmentation can effectively model emotional patterns in speech without relying on more computationally expensive deep learning frameworks. The deep Conv1D-based architecture (Fig. 15) used in the suggested model is intended to efficiently learn temporal patterns from feature sequences derived from speech. The dimensionality of feature tensors progressively decreases through successive Conv1D and max-pooling operations throughout the network. Each convolutional layer extracts higher-level temporal acoustic representations, while max-pooling layers reduce the temporal resolution of feature maps. The final convolutional feature maps are subsequently flattened into a one-dimensional vector before being passed to the fully connected dense layers for emotion classification. The tensor dimensions reported in the revised architecture diagram correspond to the actual implementation obtained from the model summary of the proposed Conv1D framework. The network can capture both higher-level temporal abstractions and low-level acoustic cues thanks to several stacked Conv1D layers with gradually decreasing temporal resolution and different filter depths. Max-pooling layers gradually down sample the feature maps to minimize overfitting and reduce computational complexity, while batch normalization layers are added after each convolution to stabilize training, speed up convergence, and lessen internal covariate shift. The learned feature maps are flattened after the convolutional blocks to create a one-dimensional representation, which is then sent to a fully connected layer to represent intricate non-linear relationships between extracted features. Multi-class classification is made possible by the final dense layer, which has seven neurons that correspond to the target emotion classes. Overall, the architecture strikes a balance between parameter efficiency and depth, producing a reliable model that can capture the discriminative temporal characteristics necessary for precise speech emotion recognition. Figure 15 illustrates the detailed architecture of the proposed deep learning framework designed specifically for speech emotion recognition. Unlike Fig. 1, which presents a general workflow of the SER pipeline, Fig. 15 focuses on the internal structure of the proposed model and the sequence of operations involved in emotion classification. The model receives the extracted acoustic feature vectors, including MFCC, ZCR, and RMSE, as input sequences that represent the temporal dynamics of speech signals. These feature sequences are processed through multiple stacked Conv1D layers that learn hierarchical temporal representations of emotional speech patterns. The first convolutional layers capture low-level acoustic cues such as pitch variations, spectral energy distribution, and short-term temporal fluctuations. Deeper convolutional layers progressively learn higher-level emotional representations by combining local acoustic patterns across longer time windows. Batch normalization layers are applied after convolutional operations to stabilize the training process and accelerate convergence by reducing internal covariate shift. Max-pooling layers are used to reduce the dimensionality of feature maps while retaining the most discriminative emotional features. The learned representations are then flattened and passed to a fully connected dense layer that models complex nonlinear relationships among extracted features. Finally, a Softmax output layer with seven neurons performs multi-class emotion classification corresponding to the seven emotional states considered in the study. This architecture enables efficient learning of temporal emotional cues from speech while maintaining a computational structure suitable for practical SER applications. Fig. 15Proposed deep conv1D-based speech emotion recognition framework. Proposed deep conv1D-based speech emotion recognition framework. The proposed architecture consists of multiple stacked Conv1D layers followed by batch normalization and max-pooling operations to progressively extract temporal emotional patterns from speech signals. These layers enable the model to capture local acoustic variations while reducing dimensionality and improving training stability. The final fully connected layer with a Softmax activation performs multi-class emotion classification across the seven emotional categories. To ensure computational efficiency, the proposed 1D-CNN architecture was designed with a relatively small number of parameters compared with deeper speech emotion recognition models. The network uses stacked convolutional layers with gradually decreasing filter sizes, enabling efficient feature extraction while maintaining a architecture. This design choice reduces computational cost and memory requirements, making the model suitable for practical applications such as real-time emotion recognition and deployment on resource-constrained systems. Conv1D layers: Conv1D layers: Conv1D layers: These layers extract temporal acoustic patterns from speech signals, capturing features such as pitch variations, spectral transitions, and short-term energy changes associated with emotional expressions. Conv1D (512 filters used for high feature diversity, which means Emotion is subtle) Low-level feature extraction explain Short-term spectral patterns like Pitch micro-variations, Energy bursts, Voice quality (breathy, tense, flat)56.4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y_{k} \left( t \right) = \sum {\:_{{i = 1}}^{K} } x\left( {t + i} \right).w_{k} \left( i \right) + \:b_{k}$$\end{document} Each filter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{k}$$\end{document}learns a distinct emotional cue. Each convolutional filter wk learns a distinct emotional cue by responding to specific temporal acoustic patterns in the input speech signal. Max pooling: Max pooling: Max pooling: Max-pooling reduces the dimensionality of feature maps while preserving the most significant emotional features, improving computational efficiency and reducing overfitting.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y\left( t \right) = {\mathrm{max}}\left( {x_{{t:t + p}} } \right)$$\end{document} Makes model invariant to small timing shifts and keeps strongest emotional evidence57. Conv1D (512 → 256 filters) Mid-level temporal abstraction is used to learn prosodic patterns (Intonation contours, Speaking rate, Stress patterns). Stacked convolutions increase effective receptive field, so deeper layers “see” longer temporal context without RNNs.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R_{{eff}} = \sum {\:_{{i = 1}}^{L} } (k_{l} - 1)$$\end{document} Fully connected layers: Fully connected layers: Fully connected layers: The dense layers combine the learned representations and model nonlinear relationships among features before the final Softmax layer performs multi-class emotion classification. Dense (512) learns non-linear combinations of emotional cues and acts as a decision manifold58.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{*{20}c} {y = \sigma \:\left( {W_{x} + b} \right)} \\ \end{array}$$\end{document} Emotion classes are not linearly separable. In Dense (7) SoftMax for 7 emotions are explained through8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P\left( {y = i} \right) = \frac{{e^{{z_{i} }} }}{{\sum \: _{j} e^{{z_{j} }} }}$$\end{document} 1D CNN matches speech physics like 1D time signal and Emotion encoded in temporal dynamics and Batch Normalization is useful for stable gradients, faster convergence, reduced covariate shift59.9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x = \frac{{x - \mu \:}}{{\sigma \:}}$$\end{document} Pseudo code for proposed work Algorithm 1Proposed speech emotion recognition framework. Proposed speech emotion recognition framework. Evaluation metrics We implement four significant evaluation metrics—accuracy, precision, recall, and F1-score—to gauge the effectiveness of the suggested Speech Emotion Recognition (SER) model60–62. These metrics offer a thorough examination of the way the model classifies various emotional categories. The pro-portion of correctly predicted observations to all observations is known as accuracy (A).10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Accuracy = \frac{{TP + TN}}{{TP + TN + FP + FN}}$$\end{document} The percentage of anticipated positive observations that are accurately classified is known as precision (P).11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Precision = \frac{{TP}}{{TP + FP}}\:$$\end{document} Recall (R) calculates the proportion of correctly identified positive observations to all class observations.12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Recall = \frac{{TP}}{{TP + FN}}$$\end{document} F1-score is the harmonic mean of precision and recall, and it ranges between 0 and 1, with 1 being the best score.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{*{20}c} {F1 - score = \frac{{2 \times \:Precision \times \:Recall}}{{Precision + Recall}}} \\ \end{array}$$\end{document} Specificity: Specificity, also referred to as the true negative rate, measures the ability of the classifier to correctly identify negative samples. It indicates how effectively the model avoids incorrectly labeling negative samples as positive63–66. The specificity metric is defined as:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{Specificity}} = \frac{{{\mathrm{TN}}}}{{{\mathrm{TN}} + {\mathrm{FP}}}}$$\end{document} A high specificity value indicates that the classifier has a strong ability to correctly reject non-target classes. In emotion recognition systems, this helps prevent misclassification of unrelated emotional states. Geometric mean (G-Mean): The Geometric Mean (G-Mean) is commonly used in classification tasks involving imbalanced datasets. It measures the balance between the classifier’s ability to correctly detect positive samples and correctly identify negative samples63–66. The G-Mean is calculated as the geometric mean of sensitivity (recall) and specificity:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{G}\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}=\sqrt{\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}\times\:\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}}\:$$\end{document} Since sensitivity is equivalent to recall, the equation can also be expressed as: 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$GMean = \sqrt {{\mathrm{Recall}} \times \:{\mathrm{Specificity}}\:}$$\end{document}A higher G-Mean value indicates that the classifier performs consistently well across both majority and minority classes, making it an important metric for evaluating classification models trained on imbalanced emotion datasets. Matthews correlation coefficient (MCC): It is considered one of the most reliable evaluation metrics for classification tasks. It takes into account all four components of the confusion matrix and provides a balanced measure even when class distributions are uneven63–66. The MCC is defined as:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MCC = \frac{{{\mathrm{TP}} \times \:{\mathrm{TN}} - {\mathrm{FP}} \times \:{\mathrm{FN}}}}{{\sqrt {\left( {{\mathrm{TP}} + {\mathrm{FP}})({\mathrm{TP}} + {\mathrm{FN}})({\mathrm{TN}} + {\mathrm{FP}})({\mathrm{TN}} + {\mathrm{FN}}} \right)} }}\:$$\end{document} The MCC value ranges between − 1 and + 1, where: +1 indicates perfect classification, 0 indicates random prediction, − 1 indicates complete disagreement between predicted and actual labels. Higher MCC values indicate a stronger correlation between predicted and actual classifications. Cohen’s kappa coefficient: Cohen’s Kappa coefficient measures the agreement between predicted and actual classifications while considering the possibility of agreement occurring by chance63–66. The Kappa coefficient is defined as:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\kappa \: = \frac{{P_{o} - P_{e} }}{{1 - P_{e} }}\:\:\:$$\end{document} where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{o}$$\end{document}represents the observed agreement between predictions and actual labels, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{e}$$\end{document}represents the expected agreement by chance, The value of the Kappa coefficient ranges from − 1 to 1, where: 1 indicates perfect agreement, 0 indicates agreement equal to random chance, negative values indicate disagreement, Values above 0.80 are generally interpreted as almost perfect agreement, indicating that the classifier predictions are highly reliable. Area under curve (AUC): The AUC represents the area under the Receiver Operating Characteristic (ROC) curve. It measures the ability of the classifier to distinguish between classes63–66. 19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$AUC = \int_{0}^{1} {TPR\left( {FPR} \right){\mathrm{d}}\left( {{\mathrm{FPR}}} \right)}$$\end{document} Where TPR = True Positive Rate, FPR = False Positive Rate. Confidence interval for accuracy: To further validate the reliability of the classification results, a 95% confidence interval (CI) for the model accuracy was computed. Confidence intervals provide a range within which the true population parameter is expected to lie with a given probability63–66. The confidence interval for accuracy is calculated as20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CI = \hat{p} \pm \:Z_{{\frac{{\alpha \:}}{2}}} \sqrt {\frac{{\hat{p}\left( {1 - \hat{p}} \right)}}{n}} \:$$\end{document} where, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{p}$$\end{document}= observed accuracy. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{\alpha\:/2}$$\end{document}= critical value of the standard normal distribution (1.96 for 95% CI). n= total number of samples. The mean value represents the average performance across all classes or evaluation runs, providing an overall measure of the model’s predictive capability. It is calculated as63–66. 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{Mean}} = \frac{1}{N}\sum {\:_{{i = 1}}^{N} } x_{i}$$\end{document} The mean value represents the average performance across all classes or evaluation runs, providing an overall measure of the model’s predictive capability. It is calculated as63–66. The mean value represents the average performance across all classes or evaluation runs, providing an overall measure of the model’s predictive capability. It is calculated as63–66. where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} denotes the evaluation value of each observation. The standard deviation (STD) evaluates the dispersion or variability of the performance metrics, indicating the stability and robustness of the model. A lower STD implies that the model produces consistent predictions across different samples. It is computed as63–66. 22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$STD = \sqrt {\frac{1}{N}\sum {\:_{{i = 1}}^{N} } (x_{i} - \mu \:)^{2} }$$\end{document} The standard deviation (STD) evaluates the dispersion or variability of the performance metrics, indicating the stability and robustness of the model. A lower STD implies that the model produces consistent predictions across different samples. It is computed as63–66. The standard deviation (STD) evaluates the dispersion or variability of the performance metrics, indicating the stability and robustness of the model. A lower STD implies that the model produces consistent predictions across different samples. It is computed as63–66. where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document} is the mean value of the observations. Results and discussion This section presents an overview of the experimental outcomes obtained from the extensive evaluation of machine learning and deep learning models for speech emotion recognition, as outlined in the preceding sections. The performance analysis is conducted across multiple benchmark datasets—RAVDESS, SAVEE, CREMA-D, and TESS—to ensure robustness and generalizability of the findings. Standard evaluation metrics, including accuracy, precision, recall, and F1-score, are employed to provide a comprehensive assessment of model effectiveness. Emphasis is placed on comparing traditional classifiers with deep learning architectures and on analyzing the impact of discriminative acoustic features and data augmentation strategies. While detailed, emotion-wise and model-specific discussions are presented later in this section, the following results collectively highlight the superiority of the proposed 1D-CNN framework, demonstrating consistent and state-of-the-art performance across diverse emotional categories and datasets. All experiments were conducted using a high-performance computing system equipped with an Intel® Core™ i9-10900 K processor, an NVIDIA RTX 3080 Ti GPU, 64 GB RAM, and a 1 TB SSD. The proposed speech emotion recognition model was implemented using Python with deep learning libraries such as TensorFlow/Keras and supporting scientific libraries including NumPy, Pandas, and Scikit-learn for data preprocessing and evaluation. Audio feature extraction was performed using standard signal processing libraries, and the experiments were executed on a Windows-based environment. This computational setup ensured efficient model training and reliable evaluation across the combined SER datasets. The reported performance metrics correspond to the evaluation of the final trained model on the held-out test dataset obtained after the train–validation–test split. The original combined dataset consisted of 12,162 samples, which were divided into training (8,756 samples), validation (973 samples), and testing (2,433 samples) subsets. Audio augmentation was applied exclusively to the training dataset, increasing the effective training samples to 11,682. All reported evaluation metrics, confusion matrices, and class-wise support values correspond only to the independent test dataset. The results are therefore presented as single evaluation values rather than mean ± standard deviation from repeated runs. The class-wise performance metrics reported in Table 2 were computed from the confusion matrix and verified against the class distribution of the test dataset. The proposed framework employs a 1D Convolutional Neural Network (1D-CNN) trained on acoustic features including Mel-Frequency Cepstral Coefficients (MFCC), Zero Crossing Rate (ZCR), and Root Mean Square Energy (RMSE). The network consists of five convolutional layers with filter sizes ranging from 128 to 512, each followed by batch normalization and max-pooling layers to stabilize training and reduce overfitting. A fully connected dense layer with 512 neurons is used before the final softmax output layer with seven emotion classes. The model was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 50 training epochs. The categorical cross-entropy loss function was used for multi-class classification. Regularization was achieved through batch normalization applied after each convolutional layer, which helps improve convergence and generalization. The complete architecture contains approximately 7.19 million parameters, of which 7.18 million are trainable. The Table 3 shows the Hyperparameter Configuration of the Proposed 1D-CNN Model. Table 3Hyperparameter configuration of the proposed model.CategoryHyperparameterValueInputInput Feature TypeMFCC + ZCR + RMSEInput Shape(2376, 1)Convolution Layer 1Filters512Kernel Size5ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 2Filters512Kernel Size5ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 3Filters256Kernel Size3ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 4Filters256Kernel Size3ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 5Filters128Kernel Size3ActivationReLUBatch NormalizationYesMax Pooling Size2Fully Connected LayerDense Units512ActivationReLUOutput LayerOutput Units7ActivationSoftmaxTraining ParametersOptimizerAdamLearning Rate0.001Batch Size32Epochs50Loss FunctionCategorical Cross-EntropyRegularizationBatch NormalizationApplied after each Conv layerTotal ParametersTrainable Parameters7,188,871Non-Trainable Parameters4,352 Table 3Hyperparameter configuration of the proposed model.CategoryHyperparameterValueInputInput Feature TypeMFCC + ZCR + RMSEInput Shape(2376, 1)Convolution Layer 1Filters512Kernel Size5ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 2Filters512Kernel Size5ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 3Filters256Kernel Size3ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 4Filters256Kernel Size3ActivationReLUBatch NormalizationYesMax Pooling Size2Convolution Layer 5Filters128Kernel Size3ActivationReLUBatch NormalizationYesMax Pooling Size2Fully Connected LayerDense Units512ActivationReLUOutput LayerOutput Units7ActivationSoftmaxTraining ParametersOptimizerAdamLearning Rate0.001Batch Size32Epochs50Loss FunctionCategorical Cross-EntropyRegularizationBatch NormalizationApplied after each Conv layerTotal ParametersTrainable Parameters7,188,871Non-Trainable Parameters4,352 Hyperparameter configuration of the proposed model. The performance evaluation Table 4; Fig. 16 offer a comprehensive evaluation of the suggested model by displaying precision, recall, F1-score and support for each emotion class. With F1-scores ranging from 0.93 to 0.96, the results show strong class-wise discrimination and consistently high performance across all emotions. The model’s ability to capture unique acoustic patterns is demonstrated by the highest F1-scores (0.96 and 0.95, respectively) for emotions like surprise and fear. The identical macro-averaged and weighted-average F1-scores (0.94) and overall accuracy of 94.91% attest to balanced performance in both majority and minority classes. These results demonstrate the suggested emotion recognition system’s resilience and capacity for generalization. It should be noted that the reported performance is obtained using a combined dataset consisting of RAVDESS, SAVEE, TESS, and CREMA-D, which introduces greater variability in speakers and recording conditions compared with studies evaluated on a single dataset. Figure 17 depicts the progression of training accuracy and validation accuracy over consecutive training epochs. Initially, training accuracy rises swiftly, indicating the model’s effective learning of distinguishing features. After around 10–15 epochs, training accuracy nears saturation close to one, suggesting the model’s convergence. Validation accuracy exhibits a similar upward pattern, showing consistent generalization to new data. Minor fluctuations in validation accuracy are noted, which are typical in stochastic optimization and reflect variability between batches. The relatively small difference between training and validation accuracy suggests minimal overfitting. Overall, the graph confirms the model’s stable convergence and strong generalization performance. Figure 18 illustrates the model’s training and validation loss over consecutive epochs. Initially, the training loss drops significantly, reflecting effective optimization and swift learning of the model’s parameters. After around 10 to 15 epochs, the training loss reaches a minimal and stable level, indicating convergence. The validation loss also generally declines, showing enhanced performance on new data. Occasional fluctuations in validation loss are noted, which are typical in stochastic training and may result from data variability. The lack of a persistent gap between training and validation loss. Although the proposed model demonstrates high accuracy and F1-score, special attention was given to mitigating potential overfitting that may arise from dataset fusion and extensive data augmentation. Combining multiple benchmark datasets such as RAVDESS, SAVEE, TESS, and CREMA-D increases the diversity of speakers, recording conditions, and emotional expressions, which helps the model learn more generalized emotional representations rather than dataset-specific patterns. In addition, augmentation techniques including noise injection, time shifting, pitch modification, and time stretching were applied carefully to simulate realistic variations in speech while preserving emotional characteristics. The training and validation accuracy and loss curves presented in Figs. 17 and 18 show a relatively small gap between training and validation performance, indicating stable convergence and limited overfitting behavior. Furthermore, the selected acoustic features—MFCC, ZCR, and RMSE—were chosen because they capture complementary aspects of speech signals, including spectral envelope characteristics, temporal signal transitions, and energy variations associated with emotional expression. MFCC features primarily represent the perceptual spectral structure of speech, while ZCR captures signal sign transitions related to speech activity, and RMSE reflects energy dynamics in the vocal signal. The combination of these features enables the model to capture both spectral and temporal emotional cues effectively. While a full ablation study and cross-corpus evaluation could provide further insights into feature contributions and generalization capability, the comparative analysis and validation behavior observed in this study indicate that the selected feature combination and dataset integration strategy contribute to robust emotion recognition performance. The class-wise evaluation metrics shown in Table 4 were derived from the confusion matrix, and the support values correspond to the number of samples belonging to each emotion class in the test dataset suggests minimal overfitting. Overall, the loss curves confirm stable training dynamics and the proposed model’s strong generalization ability. Table 4Class-wise performance evaluation of the proposed 1D-CNN model on the test dataset.EmotionPrecisionRecallF1-scoreSupportDisgust0.940.940.94394Happy0.950.930.94409Sad0.910.940.93387Neutral0.940.920.93378Fear0.940.950.95356Angry0.940.940.94390Surprise0.950.960.96119Accuracy0.94912433Macro Avg0.940.940.942433Weighted Avg0.940.940.942433 Table 4Class-wise performance evaluation of the proposed 1D-CNN model on the test dataset.EmotionPrecisionRecallF1-scoreSupportDisgust0.940.940.94394Happy0.950.930.94409Sad0.910.940.93387Neutral0.940.920.93378Fear0.940.950.95356Angry0.940.940.94390Surprise0.950.960.96119Accuracy0.94912433Macro Avg0.940.940.942433Weighted Avg0.940.940.942433 Class-wise performance evaluation of the proposed 1D-CNN model on the test dataset. Fig. 16Class-wise performance evaluation of the proposed model based on precision, recall, and F1-score. Class-wise performance evaluation of the proposed model based on precision, recall, and F1-score. Fig. 17Training and validation accuracy curves of the proposed 1D-CNN model. Training and validation accuracy curves of the proposed 1D-CNN model. Fig. 18Training and validation loss curves during model training. Training and validation loss curves during model training. Fig. 19Test performance of the proposed model showing final accuracy and loss values. Test performance of the proposed model showing final accuracy and loss values. The Fig. 19 depicts how the trained model performs on an unseen test dataset, evaluated through two metrics: test loss and test accuracy. A test loss of around 0.24 signifies that the model makes very few prediction errors on the test samples. Meanwhile, a test accuracy of approximately 94.9% indicates excellent classification ability and dependable generalization. These findings collectively imply that the model has successfully learned distinguishing features without succumbing to overfitting. In summary, the test results validate the model’s robustness and its practical utility for emotion recognition tasks. The confusion matrix in Fig. 20 provides(predicted value in horizontal and actual value is vertically present) a summary of how well the proposed emotion recognition model performs in classifying seven different emotions: disgust, happy, sad, neutral, fear, angry, and surprise. High values along the main diagonal suggest that the model accurately classifies all these emotions, indicating its overall robustness. The emotions of disgust, happy, sad, neutral, fear, and angry each show high true positive rates, demonstrating the model’s ability to effectively distinguish these emotional states. Although the surprise category has fewer samples, it still achieves a high correct classification rate, showing good generalization despite the class imbalance. Misclassifications are infrequent and mainly occur between emotions that are acoustically simi-lar, such as sad–neutral, angry–fear, and happy–neutral. These errors can be attributed to overlapping prosodic and spectral features among closely related emotions. Overall, the matrix shows that the model achieves high accuracy with minimal inter-class confusion, confirming its effectiveness for recognizing multiple classes of speech emotions. Fig. 20Confusion matrix illustrating the classification performance of the proposed 1D-CNN model across emotional categories. Confusion matrix illustrating the classification performance of the proposed 1D-CNN model across emotional categories. To ensure that the obtained classification results are reliable and not achieved by chance, a statistical evaluation of the proposed model was conducted using several statistical indicators including Mean, Standard Deviation (STD), and Root Mean Square Error (RMSE) shown in Table 5. These statistical metrics provide a more comprehensive assessment of model stability, prediction error, and consistency across different samples. Based on these statistical measures, the proposed model achieved an accuracy of 0.9491, precision of 0.94, recall of 0.94, and F1-score of 0.94, demonstrating strong classification performance. In terms of imbalance-aware evaluation, the model achieved a mean specificity of 0.9899 with a standard deviation of 0.0051, and a mean G-Mean of 0.9658 with a standard deviation of 0.0115, indicating high robustness and balanced predictive capability across classes. Furthermore, the model obtained an RMSE value of 0.6233, which reflects low prediction error. The correlation-based evaluation metrics including Matthews Correlation Coefficient (MCC) of 0.9294 and Cohen’s Kappa of 0.9292 confirm a strong agreement between predicted and actual labels. Additionally, the ROC analysis yielded an AUC of 0.9963, demonstrating excellent discriminative capability. A 95% confidence interval for accuracy (0.9351–0.9491) further confirms the statistical reliability of the proposed framework. These statistical results demonstrate that the proposed model achieves stable, reliable, and highly accurate performance, validating its effectiveness for the considered classification task, consistent with the statistical evaluation approaches reported in recent studies63–66. Table 5Overall performance evaluation of the proposed model.CategoryMetricValueClassification MetricsAccuracy0.9491Precision0.9400Recall0.9400F1-score0.9400Imbalance MetricsMean Specificity0.9899Std Specificity0.0051Mean G-Mean0.9658Std G-Mean0.0115Correlation MetricsMCC0.9294Cohen’s Kappa0.9292Error MetricsRMSE0.6233ROC AnalysisAUC0.9963Confidence IntervalAccuracy CI (95%)0.9351–0.9491 Table 5Overall performance evaluation of the proposed model.CategoryMetricValueClassification MetricsAccuracy0.9491Precision0.9400Recall0.9400F1-score0.9400Imbalance MetricsMean Specificity0.9899Std Specificity0.0051Mean G-Mean0.9658Std G-Mean0.0115Correlation MetricsMCC0.9294Cohen’s Kappa0.9292Error MetricsRMSE0.6233ROC AnalysisAUC0.9963Confidence IntervalAccuracy CI (95%)0.9351–0.9491 Overall performance evaluation of the proposed model. Comparison with state-of-the-art methods Significant performance variation between the assessed speech emotion recognition models is evident in the comparative results. Table 6Comparative performance of the proposed model with existing SER approaches.Model / studyMethodDatasetAccuracyMacro F1-scoreZhao et al.25CNN–LSTM HybridRAVDESS90.2%0.90Pepino et al.26Transformer (Wav2Vec2)Multi-dataset92.0%0.91Ahmed et al.27CNN–LSTM–GRU EnsembleMulti-dataset93.5%0.93Barhoumi et al.67Deep Learning + Data AugmentationTESS, EmoDB, RAVDESS93.3%0.93Akinpelu et al.68Vision Transformer (ViT)RAVDESS93.8%0.93Bhanbhro et al.69Attention CNN-LSTMRAVDESS94.2%0.94Tang et al.70CNN-Transformer + AttentionMELD, RAVDESS95.2%0.95Proposed Model1D-CNN + MFCC + ZCR + RMSERAVDESS + SAVEE + TESS + CREMA-D0.94910.94 Table 6Comparative performance of the proposed model with existing SER approaches.Model / studyMethodDatasetAccuracyMacro F1-scoreZhao et al.25CNN–LSTM HybridRAVDESS90.2%0.90Pepino et al.26Transformer (Wav2Vec2)Multi-dataset92.0%0.91Ahmed et al.27CNN–LSTM–GRU EnsembleMulti-dataset93.5%0.93Barhoumi et al.67Deep Learning + Data AugmentationTESS, EmoDB, RAVDESS93.3%0.93Akinpelu et al.68Vision Transformer (ViT)RAVDESS93.8%0.93Bhanbhro et al.69Attention CNN-LSTMRAVDESS94.2%0.94Tang et al.70CNN-Transformer + AttentionMELD, RAVDESS95.2%0.95Proposed Model1D-CNN + MFCC + ZCR + RMSERAVDESS + SAVEE + TESS + CREMA-D0.94910.94 Comparative performance of the proposed model with existing SER approaches. To evaluate the effectiveness of the proposed speech emotion recognition framework, the model performance was compared with recent peer-reviewed state-of-the-art SER approaches published in reputed journals and international conferences. As presented in Table 6, several recent studies have explored hybrid deep learning architectures such as CNN–LSTM models, attention-based frameworks, transformer-based architectures, and Vision Transformer (ViT)-based approaches for speech emotion recognition. These models demonstrate strong recognition performance due to their ability to capture complex spectral and temporal dependencies in speech signals; however, many of them involve higher computational complexity and require large-scale datasets and computational resources. The comparative results indicate that transformer-based and attention-guided architectures generally achieve high classification performance, with Tang et al.70 reporting the highest accuracy of 95.2% using a CNN-Transformer with multidimensional attention mechanisms. Similarly, Bhanbhro et al.69 achieved 94.2% accuracy using an attention-enhanced CNN-LSTM framework, while Akinpelu et al.68 demonstrated the effectiveness of Vision Transformers for SER tasks with an accuracy of 93.8%. Other recent approaches, including the CNN–LSTM hybrid model by Zhao et al.25, the Wav2Vec2 transformer framework proposed by Pepino et al.26, and the CNN–LSTM–GRU ensemble model by Ahmed et al.27, also reported competitive recognition performance across benchmark SER datasets. The proposed 1D-CNN framework achieves an accuracy of 94.91% with a Macro F1-score of 0.94, demonstrating competitive performance relative to recent state-of-the-art approaches while maintaining a computationally efficient architecture. Unlike several existing studies that rely on more complex transformer or hybrid recurrent architectures, the proposed model employs a comparatively Conv1D-based framework integrated with discriminative acoustic features including MFCC, ZCR, and RMSE. In addition, the proposed approach was evaluated on a combined multi-dataset setup consisting of RAVDESS, SAVEE, TESS, and CREMA-D, which introduces greater variability in speakers, recording conditions, and emotional expressions compared with single-dataset evaluations. Therefore, the achieved performance demonstrates the robustness and generalization capability of the proposed framework under more challenging evaluation conditions. The comparative analysis presented in Fig. 21 further highlights the effectiveness of the proposed model in terms of both accuracy and Macro F1-score. Although some transformer-based methods achieve slightly higher recognition performance, they typically require significantly larger model complexity and computational resources. In contrast, the proposed 1D-CNN framework provides a balanced trade-off between recognition accuracy, computational efficiency, and model simplicity, making it suitable for practical real-time speech emotion recognition applications and resource-constrained environments. Overall, the results confirm that the proposed architecture achieves reliable and competitive performance while effectively capturing temporal emotional dynamics in speech signals. Fig. 21Comparison of speech emotion recognition models with state of the art methods. Comparison of speech emotion recognition models with state of the art methods. Discussion The experimental results demonstrate that the proposed 1D-CNN framework achieves strong performance in speech emotion recognition across multiple benchmark datasets. One of the key advantages of the proposed model is its ability to effectively capture temporal acoustic patterns using stacked Conv1D layers while maintaining relatively low computational complexity compared with deeper or transformer-based architectures. The integration of discriminative acoustic features such as MFCC, ZCR, and RMSE further enhances the model’s ability to represent emotional characteristics in speech signals. Additionally, the use of multiple data augmentation techniques, including noise injection, time shifting, pitch modification, and time stretching, improves the robustness of the model by increasing data diversity and reducing the impact of class imbalance. These design choices contribute to the high classification performance achieved by the model, as reflected by the test accuracy of 94.91% and an F1-score of 0.94. Despite these advantages, some limitations remain. First, although the model performs well on the combined benchmark datasets, its performance may vary when applied to real-world speech data containing background noise, spontaneous speech patterns, or diverse linguistic variations. Second, the proposed model relies on handcrafted acoustic features rather than raw waveform learning, which may limit the ability to capture certain complex emotional cues present in speech signals. Furthermore, the study focuses primarily on controlled datasets recorded in laboratory environments, which may not fully represent real-world conversational scenarios. Future research may address these limitations by incorporating attention mechanisms, multimodal inputs such as facial expressions or physiological signals, and cross-lingual datasets to improve model generalization and real-world applicability. The combination of MFCC, ZCR, and RMSE features was selected because these features capture complementary acoustic characteristics of emotional speech signals. MFCC features represent spectral envelope information, ZCR reflects temporal signal transitions, and RMSE captures energy variations associated with emotional intensity. Preliminary experiments indicated that combining these features improves the stability and discriminative power of the model compared with using individual feature types alone. To further position the proposed approach within the broader research landscape of speech emotion recognition (SER), it is important to consider recent state-of-the-art architectures such as hybrid CNN–LSTM networks, attention-based deep learning frameworks, and transformer-based models. These architectures are capable of modeling long-term temporal dependencies and contextual relationships in speech signals and have shown promising performance in several SER studies. However, such models typically involve significantly higher computational complexity and require large-scale datasets for effective training. In contrast, the proposed 1D-CNN framework focuses on efficient temporal feature extraction while maintaining lower computational cost, making it suitable for practical deployment scenarios such as real-time emotion recognition systems and resource-constrained environments. Despite its relatively simple architecture, the proposed model achieves competitive performance with an accuracy of 94.91% and an F1-score of 0.94, demonstrating that well-designed convolutional architectures combined with discriminative acoustic features can effectively capture emotional cues in speech signals. Further insights can be obtained from the confusion matrix presented in Fig. 20, which highlights the model’s classification behavior across different emotional categories. Most emotions show strong diagonal dominance, indicating accurate predictions. However, some misclassifications are observed between acoustically similar emotional states, particularly between sad and neutral, as well as angry and fear. These errors are expected in speech emotion recognition tasks because these emotions often share overlapping prosodic characteristics such as similar pitch ranges, speech energy patterns, and spectral distributions. For example, both sad and neutral speech typically exhibit relatively low energy and slower speaking rates, making them difficult to distinguish using acoustic features alone. Similarly, angry and fear emotions may both display high arousal and elevated vocal intensity, which can lead to confusion during classification. Despite these challenges, the relatively low number of misclassifications indicates that the proposed model successfully captures discriminative emotional patterns in speech signals. Future work may further reduce these errors by incorporating attention mechanisms, multimodal emotional cues, or contextual speech modeling techniques. Despite the strong performance achieved by the proposed framework, several limitations remain in the present study. First, the proposed model primarily relies on handcrafted acoustic features such as MFCC, ZCR, and RMSE rather than end-to-end raw waveform learning, which may limit the extraction of certain complex emotional representations. Second, the experiments were conducted using controlled benchmark datasets including RAVDESS, SAVEE, TESS, and CREMA-D, which may not fully reflect real-world conversational speech conditions involving background noise, spontaneous emotional expression, and linguistic diversity. Third, although the proposed framework demonstrated good generalization across combined datasets, a dedicated cross-corpus evaluation was not performed in the current study. In addition, a detailed ablation analysis examining the individual contribution of each acoustic feature and architectural component was not included due to computational and manuscript length constraints. Furthermore, the proposed Conv1D-based framework was not compared with large-scale pretrained end-to-end transformer models under identical experimental conditions. Future research may address these limitations by incorporating multimodal emotion recognition, cross-lingual and cross-corpus evaluation, transformer-based architectures, self-supervised learning frameworks, and comprehensive feature-wise ablation studies to further improve robustness and real-world applicability. Although the present study focuses on evaluating the overall performance of the proposed SER framework, future work will include comprehensive ablation experiments to analyze the individual contributions of different components of the model, including the impact of individual augmentation techniques, comparisons between single-dataset and fused-dataset training strategies, and the effect of different feature combinations on classification performance. Conclusion and future work This study presented a one-dimensional convolutional neural network (1D-CNN) framework for speech emotion recognition using discriminative acoustic features and advanced data augmentation techniques. The proposed approach integrates MFCC, Zero-Crossing Rate (ZCR), and Root Mean Square Energy (RMSE) features to capture complementary spectral, temporal, and energy-related characteristics of emotional speech signals. In addition, multiple benchmark datasets, including RAVDESS, SAVEE, TESS, and CREMA-D, were combined to improve dataset diversity and enhance the generalization capability of the model. Experimental results demonstrate that the proposed framework achieves strong classification performance, obtaining a test accuracy of 94.91% and an F1-score of 0.94, indicating the effectiveness of the proposed feature representation and deep learning architecture for speech emotion recognition tasks. The proposed architecture achieves competitive performance while maintaining relatively low computational complexity compared with deeper SER models. Despite these promising results, several limitations remain. First, the model was evaluated primarily on controlled benchmark datasets, which may not fully represent real-world speech conditions such as background noise, spontaneous speech, or varying recording environments. Second, the proposed framework relies on handcrafted acoustic features rather than directly learning representations from raw audio signals, which may limit the ability to capture certain complex emotional patterns. Additionally, although dataset fusion improves data diversity, cross-corpus generalization across different languages and cultural contexts still requires further investigation. Although the proposed framework demonstrates strong performance on the combined multi-dataset setup, a dedicated cross-corpus evaluation using completely unseen emotional speech datasets was not conducted in the present study. Future work will focus on comprehensive cross-corpus and cross-lingual evaluations to further analyze the generalization capability and robustness of the proposed model under diverse real-world conditions. Future research will focus on addressing these limitations by exploring attention-based deep learning architectures to better capture emotionally relevant speech segments and reduce misclassification between acoustically similar emotions. Furthermore, multimodal emotion recognition approaches integrating speech with facial expressions or physiological signals will be investigated to improve recognition robustness. The deployment of the proposed framework on edge devices and embedded systems will also be explored to enable real-time emotion recognition in practical applications such as human–computer interaction, intelligent assistants, and mental health monitoring systems. Finally, cross-lingual and cross-corpus evaluations will be conducted to further assess the generalization capability of the proposed approach across diverse speech datasets.