Laryngeal disease classification using voice data: Octave-band vs. mel-frequency filters Authors: Jaemin Song, Hyunbum Kim, Yong Oh Lee Published: 2024 ABSTRACT AbstractIntroductionLaryngeal cancer diagnosis relies on specialist examinations, but non-invasive methods using voice data are emerging with artificial intelligence (AI) advancements. Mel Frequency Cepstral Coefficients (MFCCs) are widely used for voice analysis, but Octave Frequency Spectrum Energy (OFSE) may offer better accuracy in detecting subtle voice changes.Problem statementAccurate early diagnosis of laryngeal cancer through voice data is challenging with current methods like MFCC.ObjectivesThis study compares the effectiveness of MFCC and OFSE in classifying voice data into healthy, laryngeal cancer, benign mucosal disease, and vocal fold paralysis categories.MethodsVoice samples from 363 patients were analyzed using CNN models, employing MFCC and OFSE with 1/3 octave band filters. Grad-Class Activation Mapping (Grad-CAM) was used to visualize key voice features.ResultsOFSE with 1/3 octave band filters outperformed MFCC in classification accuracy, especially in multi-class classification including laryngeal cancer, benign mucosal disease, and vocal fold paralysis groups (0.9398 ± 0.0232 vs. 0.7061 ± 0.0561). Grad-CAM analysis revealed that OFSE with 1/3 octave band filters effectively distinguished laryngeal cancer from healthy voices by focusing on increased noise in the over-formant area and changes in the fundamental frequency. The analysis also highlighted that specific narrow frequency areas, particularly in vocal fold paralysis, were critical for classification, and benign mucosal diseases occasionally resembled healthy voices, making AI differentiation between benign conditions and laryngeal cancer a significant challenge.ConclusionOFSE with 1/3 octave band filters provides superior accuracy in diagnosing laryngeal diseases including laryngeal cancer, showing potential for non-invasive, AI-driven early detection.Keywords: Laryngeal disease, Voice, 1/3 octave band filter, Octave frequency spectrum energy, MFCC (mel-frequency cepstral coefficients) Introduction Background Laryngeal cancer is a malignant tumor that develops in the larynx and poses a significant health problem worldwide due to its potential severity and prevalence [1]. Early diagnosis is essential for achieving better treatment outcomes and maintaining a good quality of life for patients after treatment, which emphasizes the need for advances in diagnostic methodologies [2]. However, most diagnoses of laryngeal cancer are made after a patient visits a hospital and undergoes a laryngoscopy by a specialist, which delays the diagnosis of laryngeal cancer [3]. Laryngeal cancer presents with various symptoms such as hoarseness, globus symptom, sore throat, and swallowing difficulty [4]. Among these, hoarseness is the most common symptom, and by utilizing it, laryngeal cancer can be detected quickly and easily through a non-invasive method. However, it is challenging to diagnose early laryngeal cancer solely based on voice because the range of normal voice is very broad, influenced by factors such as sex, age, hormonal changes, and individual physical differences. Also, voice change, early symptoms of laryngeal cancer, can occur normally and be caused by many other laryngeal diseases. In addition, reliable indicators have not yet been established although there have been efforts to develop voice biomarkers that can detect laryngeal cancer using phonetic indicators. For these reasons, voice-based AI can be applied to detect laryngeal cancer. Related works Attempts to diagnose diseases through sound have been carried out in various studies. In particular, during the COVID-19 pandemic, there were numerous efforts for remote diagnosis, leading to the development of various algorithms. Albadr et al. [5] attempted to diagnose COVID-19 using Particle Swarm Optimization-Extreme Learning Machine (PSO-ELM), a type of machine learning algorithm, and achieved a maximum accuracy of 96.5 % across 13 different scenarios. Various diagnostic methods for voice disorders, as well as for COVID-19, are also being researched. Al-Dhief et al. [6] used the Naive Bayes (NB) algorithm to classify dysphonia. In their study, they applied the NB classifier to distinguish between healthy and dysphonic voices using Mel-Frequency Cepstral Coefficient (MFCC) for feature extraction with an accuracy of 81.48 %, with a sensitivity of 65 %, specificity of 91.17 %, and a G-mean of 76.98 %. The Fast Learning Network (FLN) classifier with MFCC [7] and Decision Tree (DT) algorithm based on the MFCC [8] have also been applied to classify pathologic voices. There have also been studies specifically focused on laryngeal cancer among pathologic voices. Kim et al. [9] used a one-dimensional convolutional neural network (1D-CNN) with voice data converted to MFCC to achieve an accuracy of 0.85 in distinguishing laryngeal cancer patients from healthy individuals. Similarly, Kwon et al. [10] achieved an accuracy of 0.95 for laryngeal cancer classification by combining laryngeal images and voice data through ensemble learning of two CNN models. However, when relying on voice data alone, the accuracy dropped to 0.71. Song et al. [11] improved the accuracy of laryngeal cancer classification using voice alone by collecting data with strict restrictions and optimizing parameters in the MFCC conversion. Furthermore, by embedding additional patient information and phonetic indicators that are highly correlated with laryngeal cancer into MFCC, the accuracy of laryngeal cancer classification was increased to 99 % [11]. Voice-based diagnostic methods using MFCC and CNN models for laryngeal diseases excluding laryngeal cancer have also been proposed [12,13]. These computational approaches offer a promising method for non-invasive diagnosis, potentially overcoming the limitations of human auditory diagnosis when diagnosing laryngeal cancer based on voice alone. These studies open up the possibility that artificial intelligence can detect laryngeal cancer through voice data. However, a common theme in these studies is the use of MFCCs to enable convolutional neural networks (CNN) to identify voice features that are indicative of laryngeal cancer. MFCCs are a commonly used feature set in voice processing, with frequency bands distributed on the Mel scale, which more closely resembles the response of the human auditory system than linearly spaced frequency bands. MFCCs are derived by taking the Fourier transform of a signal, mapping the powers of the spectrum to the mel scale (which approximates human hearing), and then taking the inverse Fourier transform of the logarithm of these powers. This process yields coefficients that represent the short-term power spectrum of a sound signal. As such, MFCCs are an excellent tool for capturing the essential characteristics of human voice and are widely used not only in voice recognition [14,15], but also in medical diagnostic algorithms [16] to extract, analyze, and distinguish between healthy and pathological voice signals. This paper was motivated by the proposition that octave frequency spectrum energy (OFSE) may be more suitable than traditional MFCCs for diagnosing laryngeal cancer using deep learning models. Octave band filters are a type of frequency filter that divide the audible frequency range into bands, where each band spans one octave, meaning the upper frequency limit is twice the lower frequency limit. This results in a logarithmic frequency resolution that closely aligns with the human auditory system's perception of sound, providing finer resolution at lower frequencies. The primary advantage of octave band filters, especially 1/3 octave band filters, lies in their ability to provide a non-linear frequency resolution that offers detailed analysis of audio signals, especially at lower frequencies. This characteristic is crucial for applications such as acoustic analysis, audio engineering, environmental noise monitoring, speech and voice analysis, and vibration analysis. For example, studies performed feature extraction on bearing noise using an octave filter bank and applied this to a machine learning algorithm for faulty condition classification [17,18]. Octave band filters excel in capturing the detailed spectral characteristics of complex audio signals, making them valuable for identifying subtle changes in voice patterns, which is particularly useful in medical diagnostics like laryngeal cancer detection. On the other hand, MFCCs are widely used in speech and audio processing because they effectively capture the phonetically relevant characteristics of audio signals [19]. However, they may not always capture the detailed spectral nuances that octave band filters can, especially in lower frequency ranges. Therefore, applying octave band filters can be advantageous in scenarios where precise frequency analysis is required. For instance, in the diagnosis of laryngeal cancer, octave band filters can more accurately capture the subtle voice changes associated with the disease, potentially leading to better diagnostic accuracy compared to MFCCs. By leveraging the detailed frequency resolution provided by octave band filters, researchers can improve the performance of machine learning models designed to detect laryngeal cancer, thereby offering a more reliable method for early and non-invasive diagnosis. These computational approaches offer a promising method for non-invasive diagnosis, potentially overcoming the limitations of human auditory diagnosis when diagnosing laryngeal cancer based on voice alone. This paper conducts a comparative analysis of the effects of MFCC and OFSE with 1/3 octave band filter transformations in the machine learning-based analysis of voice datasets for laryngeal cancer. First, we evaluate the classification performance across three classes—healthy, laryngeal cancer, and other laryngeal conditions demonstrating the effectiveness of each transformation method in distinguishing these categories. Second, we illustrate the differences in how MFCC and OFSE with 1/3 octave band filter recognize laryngeal cancer, providing insights into the unique characteristics captured by each method. Finally, we employ Class Activation Maps (CAMs) to visually represent the classification results, highlighting the distinctive capabilities of MFCC and OFSE with 1/3 octave band filter in identifying and diagnosing laryngeal cancer through voice data. The rest of this paper is organized as follows. Section 2 describes the methodology, including details on study participants, data conversion techniques, and the classification models used. Section 3 presents the results, with a focus on classification performance and Grad-CAM analysis for visualizing model outputs. Section 4 provides a detailed discussion of the findings, and Section 5 concludes the paper. Methods Study participants details Voice samples from 363 patients were collected for this study. It is a retrospective review method and is based on data from January 2015 to December 2022 at a single center, Yeouido St. Mary's Hospital, Catholic University of Korea, Yeouido. (This study was approved by the Institutional Review Board (IRB) of Yeouido St. Mary's Hospital, titled "Development of an artificial intelligence platform for early diagnosis of laryngeal cancer by combining voice and image", approval number XC23RIDI0042, approval date July 25, 2023). Patient selection was filtered with the following conditions: (1) voice changes were maintained for more than 3 weeks after the first visit; (2) patients under 19 years of age, pregnant women, and patients who had undergone vocal cord surgery or radiation therapy were excluded; and (3) only male patients were included to account for differences in the occurrence of laryngeal disease between male and female. /a/vocalization samples were recorded in the environment on a personal computer using a Kay Computer Speech Lab (CSL) (model 5121; KayPentax, Lincoln Park, NJ, USA), including a Shure-Prolog SM48 microphone and a Shure digital amplifier, positioned at a distance of 10–15 cm from the mouth and at a 90-degree angle. Environment noise was controlled below 45 dB HL for recording quality. Participants produced the vowel sound/a/for at least 4 s at a comfortable loudness of approximately 55–65 dB HL. A summary of related works on laryngeal cancer detection based on voice is presented in Table 1. Table 1Summary of related works on laryngeal cancer detection based on voice.Table 1ReferenceDatasetFeature ExtractionModelResultsWeaknessesKim et al. [9]Voice data of laryngeal cancer patientsMel-Frequency Cepstral Coefficients (MFCC)1D-CNNAccuracy: 0.85Limited to voice data, no additional medical info usedKwon et al. [10]Laryngeal images and voice dataMFCC + image featuresEnsemble of two CNN modelsAccuracy: 0.95 (0.71 with voice)Accuracy drops significantly when using voice aloneSong et al. [11]Voice data of laryngeal cancer patientsMFCC + phonetic indicatorsCNNAccuracy: 0.99Requires strict data collection conditions Summary of related works on laryngeal cancer detection based on voice. For healthy voice samples, they were collected during preoperative voice evaluation before thyroid or lungs surgery and confirmed to be free of lesions by laryngoscopy performed by two otolaryngologists. Larynx-related disease samples were categorized as laryngeal cancer, benign mucosal disease, and vocal cord paralysis. Since 90 % of laryngeal cancer occurs in men, only male voices were collected to eliminate sex differences. Laryngeal cancer includes glottis cancer and transglottis cancer, while benign mucosal disease includes several conditions such as vocal cord polyps, vocal cord nodules, granulomas, and Reinke's edema. Vocal fold paralysis includes decreased motility of unilateral or bilateral vocal folds. As a result, 155 healthy voice samples, 30 laryngeal cancer, 81 benign mucosal disease, and 97 vocal fold paralysis samples were collected. The detailed characteristics of the collected data are shown in Table 2. Table 2Clinical characteristics of the 363 male participants.Table 2CharacteristicsN (%)a or Mean ± SDp-valueLaryngealCancer (N = 30)Vocal fold paralysis (N = 97)Benign mucosal disease (N = 81)Healthy (N = 155)Age(years)66 ± 1155 ± 1854 ± 1446 ± 13<0.01Smoking(%)< 30pyrsb6 (20)75 (77)68 (84)144 (92)<0.01≥ 30pyrs24 (80)22 (23)13 (16)11 (7)LesionUnilateral23 (77)96 (99)54 (67)Bilateral7 (23)1 (1)27 (33)Anterior commissure involvementYes17 (57)18 (22)No13 (43)63(78)Malignancy primary siteGlottis25 (83)Transglottis5 (17)Malignancy T stagecEarly stage26 (87)Advanced stage4 (13)aPercentage of the each of the groups.bpyrs, an abbreviation for pack years, denotes the consumption of one pack of cigarettes daily over a period of one year.c8th Edition of the AJCC TNM classification (2016). Clinical characteristics of the 363 male participants. Percentage of the each of the groups. pyrs, an abbreviation for pack years, denotes the consumption of one pack of cigarettes daily over a period of one year. 8th Edition of the AJCC TNM classification (2016). Data conversion: MFCC vs. OFSE We delve into the technical nuances of Mel Frequency Cepstral Coefficients (MFCC) and octave frequency spectrum energy (OFSE), comparing their suitability and effectiveness in processing voice data for the detection of laryngeal cancer (Fig. 1). Examples of conversions for waveform, MFCC, and OFSE. Mel Frequency Cepstral Coefficients (MFCC) MFCC is a widely used feature extraction technique in the field of speech and audio processing. The process begins with the division of the audio signal into short frames, as the frequency content of audio signals varies over time. Each frame undergoes a Fourier transform to convert it from the time domain to the frequency domain. The power spectrum is then mapped onto the mel scale, which is a perceptual scale of pitches judged by listeners to be equal in distance from one another. This transformation mimics the human ear's inner workings, which do not perceive pitches linearly but logarithmically. The log of the powers at each of the mel frequencies is taken, and a discrete cosine transform (DCT) is applied [20]. This step is crucial as it converts the log mel spectrum into time coefficients, which are the MFCCs. The MFCCs succinctly capture the power distribution over the phone spectrum and are therefore highly useful in audio pattern recognition tasks. To extract the MFCC, the entire audio was segmented into 0.5-s intervals, and each segment was subsequently transformed. The analysis window length was set to 0.02 s, the step between successive windows to 0.01 s, the number of cepstral coefficients to 13, and the number of filters in the filter bank to 26 (Table 3). Table 3Frequency band details for Mel frequency filter (Hz).Table 3Filter IndexLower edgeCenter FrequencyUpper edgeFilter IndexLower edgeCenter FrequencyUpper edge1096.33205.92143041.473556.354142.09296.33205.92330.59153556.354142.094808.443205.92330.59472.41164142.094808.445566.494330.59472.41633.76174808.445566.496428.865472.41633.76817.3185566.496428.867409.916633.76817.31026.11196428.867409.918525.967817.31026.111263.65207409.918525.969795.681026.111263.651533.88218525.969795.611239.9691263.651533.881841.29229795.611239.9612883.09101533.881841.292191.022311239.9612883.0914752.34111841.292191.022588.872412883.0914752.3416878.83122191.022588.873041.472514752.3416878.8319297.96132588.873041.473556.352616878.8319297.9622050 Frequency band details for Mel frequency filter (Hz). Octave frequency spectrum energy (OFSE) In contrast to MFCC, OFSE applies an octave-band filter to the frequency bands, preserving more information in the lower frequencies which are often critical for pathological voice analysis. OFSE starts similarly by segmenting the audio signal into frames and converting these frames into the frequency domain. However, instead of mapping to the mel scale, OFSE applies an octave filter bank, which segments the power spectrum using logarithmic frequency spacing. The detailed OFSE transformation process is shown in Fig. 2. This method enhances the resolution of the frequency domain representation, especially at lower frequencies, making it potentially more effective in detecting the subtle anomalies characteristic of laryngeal pathologies. An octave filter bank is an array of octave bandpass filters that are defined based on ANSI s1.11-2004 standard and IEC 61260-1-2014 standard. In contrast to the ANSI s1.11-2004 standard, the IEC 61260-1-2014 standard requires the starting frequency of the frequency array to be 16 Hz for 1/1 octave. OFSE requires adherence to standards like ANSI s1.11-2004 and IEC 61260-1-2014 for precise octave-band filter configuration. The audio data was segmented into 0.5-s segments, and OFSE with 1/3 octave band filter was applied to each segment, with the frequency limitation set between 12 Hz and 40000 Hz (Table 4). The computational complexity of OFSE is comparable to MFCC, as both methods involve segmenting the audio into frames and transforming these frames into the frequency domain. Transformation process of octave frequency spectrum energy. Table 4Frequency band details for octave frequency spectrum energy with 1/3 octave band filter (Hz).Table 4Filter IndexLower edgeCenter FrequencyUpper edgeFilter IndexLower edgeCenter FrequencyUpper edge011.2212.5914.1318707.95794.33891.25114.1315.8517.7819891.251000.001122.02217.7819.9522.39201122.021258.931412.54322.3925.1228.18211412.541584.891778.28428.1831.6235.48221778.281995.262238.72535.4839.8144.67232238.722511.892818.38644.6750.1256.23242818.383162.283548.13756.2363.1070.79253548.133981.074466.84870.7949.4389.13264466.845011.875623.41989.13100.00112.20275623.416309.577079.4610112.20125.89141.25287079.467943.288912.5111141.25158.49177.83298912.5110000.0011220.1812177.83199.53223.873011220.1812589.2514125.3813223.87251.19281.843114125.3815848.9317782.7914281.84316.23354.813217782.7919952.6222387.2115354.81398.11446.683322387.2125118.8628183.8316446.68501.19562.343428183.8331622.7835481.3417562.34630.96707.953535481.3439810.7244668.36 Frequency band details for octave frequency spectrum energy with 1/3 octave band filter (Hz). Classification models Both MFCC and OFSE processed signals are fed into a Convolutional Neural Network (CNN) for classification. For this study, we specifically employ the ResNet architecture, known for its superior performance in various image and signal processing domains, including laryngeal cancer detection. Residual Network (ResNet): ResNet is a type of CNN that introduces the concept of residual learning to address the vanishing gradients problem that often occurs with networks as they grow deeper. Traditional deep networks struggle to train effectively due to the degradation problem; as the network depth increases, accuracy gets saturated and then degrades rapidly. ResNet addresses this by using shortcut connections, or skip connections, that perform identity mapping, allowing the network to skip one or more layers. These connections add the input of the current layer to the output of a layer deeper in the network. This approach allows the network to learn residual functions with reference to the layer inputs, making it easier to train deeper networks without a significant increase in training error. In the CNN model, several parameters were set to optimize performance. The batch size was configured to 8, while the input image size was set to (224, 224) pixels. The training process was run for 100 epochs, with the model configured to save at the point when the validation loss reached its minimum value. The architecture of ResNet comprises several stacked residual blocks, each containing two or three layers. By leveraging these residual blocks, ResNet can effectively learn from a vast amount of data and deeper architectural settings without the burden of training complications inherent in traditional deep convolutional networks. For our purpose, the application of ResNet enables the model to capture and utilize complex patterns within the cepstral features derived from the voice data, which are essential for accurately distinguishing between normal and pathological laryngeal conditions. This methodology significantly boosts the diagnostic accuracy and robustness of the model when applied to laryngeal cancer detection from voice signals. To interpret the decision-making process of the CNN, Class Activation Mapping (CAM) is used. CAM helps in visualizing the regions of the input that are important for making a prediction by the CNN. This technique generates a heatmap for each class prediction, highlighting the significant areas that contribute to the classification decision. Among various CAM method, we have applied Grad-CAM so that we can visually compare how MFCC and OFCC influence the CNN's focus areas and recognition patterns in distinguishing between normal, cancerous, and other pathological laryngeal conditions. Quantification and statistical analysis All clinical characteristics were statistically analyzed using the Mann–Whitney U test or chi-square test under Python (version 3.10.12; https://www.python.org/). A P-value of <0.05 was considered statistically significant. The ROC curves plotting and area under curve calculation were performed using the “pROC” package. ROC curves were compared using the DeLong test. Results Classification performance ResNet [21] model was trained to classify voice samples into healthy, vocal fold paralysis, benign mucosal disease in larynx, and laryngeal cancer categories. Both results using OFSE with a 1/3 octave band filter and MFCC with a Mel-filter were compared after training and testing with ResNet. Performance was evaluated using 5-fold cross-validation (CV) to ensure robustness and generalizability [22]. The following metrics were used for evaluation: accuracy, precision, recall, and F1-score. The assessment of the model's performance is based on these key metrics. True positives (TP) refers to correctly predicted positive instances, true negatives (TN) refers to correctly predicted negative instances, false positives (FP) refers to negative instances incorrectly predicted as positive, and false negatives (FN) refers to positive instances incorrectly predicted as negative. Accuracy measures the proportion of correct predictions, both positive and negative. Precision indicates the proportion of true positive predictions out of all positive predictions. Recall reflects the model's ability to identify actual positive cases. F1-Score balances precision and recall as their harmonic mean. These metrics together provide a comprehensive view of the model's classification effectiveness. The experiment was conducted using Python version 3.10.12, TensorFlow version 2.17.0, and an NVIDIA Tesla T4 GPU, which has 16 GB of memory. Source codes are available at https://github.com/jaemin-song876/Laryngeal-Disease-Classification-using-Voice-Data/tree/main. Binary classification of healthy and laryngeal cancer group In the task of distinguishing healthy voice from laryngeal cancer cases, the OFSE method demonstrated superior performance. As shown in Fig. 3, the accuracy for OFSE was 1.0000, slightly higher than MFCC's 0.9899. The precision for OFSE was 1.0000, compared to 0.9845 for MFCC. OFC's recall was 1.0000, notably better than MFCC's 0.9758. The F1-score for OFSE was 1.0000, while MFCC achieved 0.9798. Although the performance improvement is slight due to the already high accuracy of MFCC in healthy and laryngeal cancer classification, OFSE achieved 100 % accuracy in CV. Fig. 4 shows the accumulated results of the 5-fold cross-validation. Result of binary classification between healthy and laryngeal cancer groups using OFSE with a 1/3 octave band filter and MFCC Methods (mean ± SD) ∗Abbreviation: OFSE: octave frequency spectrum energy, MFCC: mel-frequency cepstral coefficients. Confusion matrix for classification among healthy, laryngeal cancer groups. Binary classification of healthy and vocal fold paralysis group In the task of distinguishing healthy from vocal fold paralysis cases, the OFSE method again demonstrated superior performance. As shown in Fig. 5, the accuracy for OFSE was 0.9884, significantly higher than MFCC's 0.9047. The precision for OFSE was 0.9877, compared to 0.9018 for MFCC. OFSE's recall was 0.9876, notably better than MFCC's 0.8949. The F1-score for OFSE was 0.9876, while MFCC achieved 0.8978. These results indicate substantial improvements across all metrics when using the OFSE method. Additionally, the standard deviations for OFSE were lower, further confirming its consistency and robustness. Fig. 6 shows the accumulated results of the 5-fold cross-validation. Result of binary classification between healthy and vocal fold paralysis groups using OFSE with a 1/3 octave band filter and MFCC Methods (mean ± SD) ∗Abbreviation: OFSE: octave frequency spectrum energy, MFCC: mel-frequency cepstral coefficients. Confusion matrix for classification among healthy and vocal fold paralysis groups. Binary classification of healthy and benign mucosal disease in larynx group For the binary classification of healthy versus benign mucosal disease in larynx cases, the OFSE method continued to outperform the MFCC method. As shown in Fig. 7, the accuracy for OFSE was 0.9855, higher than MFCC's 0.8995. OFSE achieved a precision of 0.9866, surpassing MFCC's 0.8982. The recall for OFSE was 0.9819, compared to 0.8810 for MFCC. Consequently, the F1-score for OFSE was 0.9841, higher than MFCC's 0.8882. These results further highlight the enhanced performance of the OFSE preprocessing method. Moreover, the standard deviations for OFSE were significantly lower, indicating better reliability in performance. In addition, it is observed that the accuracy for benign disease was significantly lower when using MFCC compared to other conditions such as cancer and paralysis. However, there was an improvement in accuracy with the use of OFSE. This indicates that OFSE processed benign features more effectively than MFCC. Fig. 8 presents the cumulative results obtained from the 5-fold cross-validation. Result of binary classification between healthy and benign mucosal disease in larynx groups using OFSE with a 1/3 octave band filter and MFCC Methods (mean ± SD) ∗Abbreviation: OFSE: octave frequency spectrum energy, MFCC: mel-frequency cepstral coefficients. Confusion matrix for classification among healthy and benign mucosal disease in larynx groups. As shown in Fig. 9, the receiver operating characteristics (ROC) curves for binary classifications of healthy-laryngeal cancer, healthy-vocal fold paralysis, and healthy-benign mucosal disease all exhibit very high area under ROC (AUROC) values of 0.99. These results indicate that the OFSE method with CNN achieves excellent discrimination capability across these classifications. The consistently high AUROC scores across all binary classifications further emphasize the robustness and effectiveness of the OFSE approach in capturing critical features for accurate diagnosis. Receiver operating characteristics (ROC) curves for binary models using OFSE with a 1/3 octave band filter ∗Abbreviation: OFSE: octave frequency spectrum energy. Four-class classification of healthy, laryngeal cancer, vocal fold paralysis, and benign mucosal disease in larynx group In the comprehensive task of classifying healthy, laryngeal cancer, vocal fold paralysis, and benign mucosal disease cases, the OFSE method showed significant superiority over the MFCC method. As shown in Fig. 10, the accuracy for OFSE was 0.9398, whereas MFCC achieved 0.7061. The precision for OFSE was 0.9417, compared to 0.7186 for MFCC. OFSE's recall was 0.9232, while MFCC's recall was 0.6303. The F1-score for OFSE was 0.9314, significantly better than MFCC's 0.6438. These results indicate that OFSE preprocessing provides a substantial improvement in the classification performance across all metrics. The standard deviations for all metrics were also lower for the OFSE method, demonstrating more consistent and reliable performance. More details about classification results in Fig. 11. Result of classification among healthy, laryngeal cancer, vocal fold paralysis, and benign mucosal disease in larynx groups using OFSE with a 1/3 octave band filter and MFCC Methods (mean ± SD) ∗Abbreviation: OFSE: octave frequency spectrum energy, MFCC: mel-frequency cepstral coefficients. Confusion matrix for classification among healthy, laryngeal cancer, vocal fold paralysis, and benign mucosal disease in larynx groups. Grad-CAM analysis To interpret the decision-making process of the CNN with OFSE, Class Activation Mapping (CAM) was used [23,24]. Grad-CAM helps in visualizing the regions of the input that are important for making a prediction by the CNN. This technique generates a heatmap for each class prediction, highlighting the significant areas that contribute to the classification decision. Among various CAM methods, we have applied Grad-CAM to visually compare the distribution of clusters for different classes within OFSE [25]. This allows us to identify specific frequency bands within the OFSE that are important for discriminating laryngeal diseases. OFSE, especially 1/3 octave band filters, divided the frequency range from 11 Hz to 44,668 Hz into 36 segments, each covering 0–35 units. The fundamental frequency range of 90–130 Hz [[26], [27], [28]] falls within the 9th and 10th segments, and this range is designated as the fundamental area. Additionally, the 1st and 2nd formant frequency range of 700–1500 Hz [[26], [27], [28]] spans the 17th to 21st segments, and this range is designated as the formant area. Occasionally, the 3rd formant appeared in the 23rd to 25th segments, but there were instances where it did not appear. For this reason, the 22nd to 35th segments were designated as the over-formant area. In conclusion, the frequency band ranges were categorized as follows: an index of 8 or below was designated as under-fundamental area, 9 to 10 as fundamental area, 11 to 16 as under-formant area, 17 to 21 as formant area, and above 21 as over-formant area (Table 5). Table 5Frequency band categories according to frequency ranges.Table 5CategoryFrequency range (Hz)Segment of OFSEUnder-fundamental area11–991–8Fundamental area99–1419–10Under-formant area141–56211–16Formant area562–177817–21Over-formant area1778–4466822–35∗Abbreviation: OFSE: octave frequency spectrum energy. Frequency band categories according to frequency ranges. ∗Abbreviation: OFSE: octave frequency spectrum energy. According to the source-filter theory [29], the fundamental frequency is generated in the larynx, while filtering and resonance occur in the vocal tract, which includes the pharynx, oral cavity, and nasal cavity. Therefore, for phonetic changes related to laryngeal disease, it is more accurate to primarily analyze the fundamental frequency, with formant analysis serving as a supplementary measure. MFCC divides frequencies under 130 Hz into only 2 index filters, whereas OFSE divides this range into 10 index filters, allowing for more refined classification. Additionally, MFCC divides frequencies below 700 Hz, the lower limit of the 1st formant, into 6 index filters, while OFSE uses 17 index filters. Therefore, OFSE, which categorizes the under-formant range with more filters, can represent acoustic changes due to laryngeal disease more subtly. The first and second formant area is divided into 5 index filters by both MFCC and OFSE. In conclusion, OFSE divides the fundamental frequency range more precisely than MFCC and classifies the formant range with similar accuracy to MFCC. In addition, high-frequency band signal characteristics beyond index 21 may also be important in diagnosing laryngeal cancer. These high-frequency components are more closely associated with the timbre and intricate details of the voice signal. The ability to identify voice texture and clarity, which are often impaired in laryngeal cancer patients, is heavily influenced by these high-frequency components. While detailed phonetic and pathological analysis of the phonetic characteristics of laryngeal disease has been limited, the following analysis aims to fill this gap. As shown in the previous experiments, CNN model classification based on OFSE with 1/3 octave band filter has high discriminative capability. Through the analysis of the critical activated regions of these CNN models, we aim to systematically investigate the importance of different frequency bands in identifying laryngeal diseases, and as mentioned earlier, this approach not only highlights the important role of the fundamental and formant area, but also emphasizes the complementary value of higher frequency bands. Analysis of the Grad-CAM patterns To facilitate pattern analysis, we performed clustering on the Grad-CAM results. Given the current OFSE size of 228x228, directly applying k-means clustering would be computationally intensive. Therefore, we used average pooling with an 8x8 kernel size and stride of 8 to reduce the size of the Grad-CAM output, resulting in an abstracted Grad-CAM with a reduced resolution. This abstracted Grad-CAM was then subjected to k-means clustering to identify and analyze patterns in Fig. 12. Abstracting Grad-CAM for clustering. Healthy and laryngeal cancer binary classification We utilized Grad-CAM images derived from the binary classification of laryngeal cancer versus healthy conditions to analyze the regions of interest within OFSE with 1/3 octave band filter. To effectively classify the Grad- CAMs, we used the K-means clustering method. For the binary classification of laryngeal cancer and healthy voice, we set the k value to 8 for clustering. If a cluster encompassed multiple frequency ranges, it was named based on the most dominant frequency within that cluster. The detail feature is shown in Fig. 13, and the proportion of clusters within each class for each cluster in laryngeal cancer and healthy voice is shown in Fig. 14. •Early Broad formant area highlighting: Activation in this pattern is prominent in the formant area but occurs mainly in the early time frames. Early Broad formant area highlighting: Activation in this pattern is prominent in the formant area but occurs mainly in the early time frames. •Broad formant and over-formant area highlighting: The activation covers a broad range of formant area in the early time frame, and subsequently includes over-formant area, particularly the 23rd to 25th segments corresponding to the 3rd formant, as illustrated in Fig. 15. After early time frame, the activation situated in over-formant area mainly. Broad formant and over-formant area highlighting: The activation covers a broad range of formant area in the early time frame, and subsequently includes over-formant area, particularly the 23rd to 25th segments corresponding to the 3rd formant, as illustrated in Fig. 15. After early time frame, the activation situated in over-formant area mainly. •Broad over-formant area highlighting: This pattern shows multiple specific areas of activation within the over-formant area across the entire time frame, indicating that over formant frequency, including third formant, has difference between laryngeal cancer and healthy conditions. Broad over-formant area highlighting: This pattern shows multiple specific areas of activation within the over-formant area across the entire time frame, indicating that over formant frequency, including third formant, has difference between laryngeal cancer and healthy conditions. •Narrow over-formant area highlighting: This pattern indicates a narrow activation spanning than broad over-formant area in the entire time frame, serving as a distinguishing feature between laryngeal cancer and healthy voice. Narrow over-formant area highlighting: This pattern indicates a narrow activation spanning than broad over-formant area in the entire time frame, serving as a distinguishing feature between laryngeal cancer and healthy voice. •Extremely over-formant area highlighting: This pattern exhibits prominent activation in the extremely over-formant area, primarily occurring in the later time frames. Extremely over-formant area highlighting: This pattern exhibits prominent activation in the extremely over-formant area, primarily occurring in the later time frames. •Late formant & under-formant area highlighting: This pattern exhibits prominent activation in the formant and under-formant area, primarily occurring in the later time frames. Late formant & under-formant area highlighting: This pattern exhibits prominent activation in the formant and under-formant area, primarily occurring in the later time frames. •Fundamental & under-fundamental area highlighting: This pattern demonstrates significant activation in the fundamental and under-fundamental area, predominantly occurring in the later time frames. Fundamental & under-fundamental area highlighting: This pattern demonstrates significant activation in the fundamental and under-fundamental area, predominantly occurring in the later time frames. •Narrow under-fundamental area highlighting: This pattern is distinguished by significant activation within the specific under-fundamental area, highlighting that low-frequency components, play a crucial role in the classification decisions for certain cases. Narrow under-fundamental area highlighting: This pattern is distinguished by significant activation within the specific under-fundamental area, highlighting that low-frequency components, play a crucial role in the classification decisions for certain cases. K-means Clustering of Grad-CAM Images in binary classification between laryngeal cancer and healthy voice groups. Distribution values for each cluster in laryngeal cancer and healthy voice groups ∗Abbreviation: NOF: Narrow over-formant area highlighting, BOF: Broad over-formant area highlighting, NUF: Narrow under-fundamental area highlighting, F&UF: Fundamental & under-fundamental area highlighting, BF&OF: Broad formant and over-formant area highlighting, LF&UF: Late formant & under-formant area highlighting, EOF: Extremely over-formant area highlighting, EBF: Early Broad formant area highlighting. Distribution values for each two voice groups when dividing clusters into three classifications: over-formant, formant and under-formant. According to Grad-CAM and K-means in Fig. 13, all frequency energy from the under-fundamental frequency to the over-formant frequency was used to recognize the pathologic voice. Formant area was main factor in three clusters: Early broad formant, broad formant and over-formant, and late formant & under-formant area highlighting. Under-fundamental and fundamental area was main factor in two clusters: Fundamental & under-fundamental, and narrow under-fundamental area highlighting. This means that the difference between cancer and healthy voice within the formant area and the area around the fundamental frequency, as well as in the over-formant area, plays an important role in distinguishing laryngeal cancer. In Fig. 14, laryngeal cancer was found in only four clusters: Broad over-formant, narrow over-formant, narrow under-fundamental, and fundamental & under-fundamental area highlighting. According to these results, the difference between laryngeal cancer and healthy voice was distinguished based on unique characteristics found in the under-fundamental area and the over-formant area, rather than the formant area. In terms of proportion in Fig. 13, about 28 % of laryngeal cancer cases were diagnosed through the under-fundamental area. Of the 72 % of laryngeal cancer cases diagnosed in the over-formant area. Overall, in the binary classification of laryngeal cancer and healthy voices, the over-formant area was used in 48 % of the cases, the formant area in 42 %, and the fundamental area, including the under-formant area, in 34 %, including overlaps. The difference between early and late activation times influenced the cluster classification in Fig. 13: Early broad formant (first 0.05 s), and late formant & under-formant area highlighting (last 0.05 s). However, this had no effect on the pathologic differences in Fig. 12. This means that while laryngeal cancer exhibits abnormalities throughout the entire time range, a healthy voice can be distinguished if there is even a very short segment without abnormalities. Multi-class classification: healthy, laryngeal cancer, vocal fold paralysis, and benign mucosal disease in larynx group In the 4-classes classification, we employed k-means clustering with k set to 10. The following example clusters illustrate the images, and the detail feature is shown in Fig. 16. The distribution values for each cluster in multi-class classification is shown in Fig. 17. •Broad formant area highlighting: The activation covers a broad range of formant area in the early time frame, and subsequently includes over-formant area, particularly the 23rd to 25th segments corresponding to the 3rd formant. Broad formant area highlighting: The activation covers a broad range of formant area in the early time frame, and subsequently includes over-formant area, particularly the 23rd to 25th segments corresponding to the 3rd formant. •Narrow formant area highlighting: The activation spans in range of formant area throughout the entire time frame. Narrow formant area highlighting: The activation spans in range of formant area throughout the entire time frame. •Early narrow formant area highlighting: Activation in this pattern is prominent in the formant area but occurs mainly in the early time frames. Early narrow formant area highlighting: Activation in this pattern is prominent in the formant area but occurs mainly in the early time frames. •Broad over-formant area highlighting: The activation covers a broad range of over-formant area across the entire time frame. This activation sometimes includes formant area, however, main activation is located at over-formant area. Broad over-formant area highlighting: The activation covers a broad range of over-formant area across the entire time frame. This activation sometimes includes formant area, however, main activation is located at over-formant area. •Narrow over-formant area highlighting: The activation covers only range of over-formant area across the entire time frame. Narrow over-formant area highlighting: The activation covers only range of over-formant area across the entire time frame. •Late formant & over-formant area highlighting: Activation in this pattern is prominent in formant and over-formant area but occurs mainly in the later time frames. Late formant & over-formant area highlighting: Activation in this pattern is prominent in formant and over-formant area but occurs mainly in the later time frames. •Under-formant area highlighting: This pattern shows multiple specific areas of activation within the under-formant area across the entire time frame. This activation sometimes includes fundamental area, however, main activation is located at under-formant area. Under-formant area highlighting: This pattern shows multiple specific areas of activation within the under-formant area across the entire time frame. This activation sometimes includes fundamental area, however, main activation is located at under-formant area. •Multiple under-fundamental area highlighting: This pattern is characterized by activation in the under-fundamental area. It signifies that low-frequency components of the voice signal are most influential in the classification decision for certain cases. Multiple under-fundamental area highlighting: This pattern is characterized by activation in the under-fundamental area. It signifies that low-frequency components of the voice signal are most influential in the classification decision for certain cases. •Single under-formant area highlighting: This pattern is characterized by activation within specific frequency bands, especially under-formant area, within 0.05 s. Single under-formant area highlighting: This pattern is characterized by activation within specific frequency bands, especially under-formant area, within 0.05 s. •No area highlighting: Activation in this pattern is prominent in the Low-frequency range but occurs mainly in the later time frames. No area highlighting: Activation in this pattern is prominent in the Low-frequency range but occurs mainly in the later time frames. K-means Clustering of Grad-CAM Images in binary classification among laryngeal cancer, vocal fold paralysis, benign mucosal disease in larynx, and healthy voice groups. Distribution values for each cluster in laryngeal cancer, vocal fold paralysis, benign mucosal disease in larynx, and healthy voice groups ∗Abbreviation: BF: Broad formant area highlighting, UF: Under-formant area highlighting, NOF: Narrow over-formant area highlighting, NF: Narrow formant area highlighting, MUF: Multiple under-fundamental area highlighting, LF&OF: Late formant & over-formant area highlighting, ENF: Early narrow formant area highlighting, SUF: Single under-formant area highlighting, BOF: Broad over-formant area highlighting, NAE: No area highlighting. In the multi-class classification, more clusters were generated. A total of 10 clusters were created, including a cluster without enhancement: no area highlighting. According to Fig. 18, no area enhancement is observed only in the benign mucosal disease and healthy groups. A total of 43 cases (86 %) of the voices without enhancement were normal, while 7 cases (14 %) were benign. There were no cases for laryngeal cancer and vocal fold paralysis in no area highlighting, indicating that both laryngeal cancer and vocal fold paralysis always exhibit some characteristic sound even within a short duration of 0.5 s. Distribution values for each four voice groups when dividing clusters into three classifications: over-formant, formant and under-formant. Unlike in binary classification, where the over-formant area is used the most, in multi-class classification, the formant area was used the most at 57 %. The fundamental area, including the under-formant area, was used in 34 % of cases in binary classification but decreased to 24 % in multi-class classification. However, the proportion of classifications using only the under-formant area increased from 19 % in binary classification to 24 %. Contrary to frequency variations, regardless of early or late activation, a short activation time indicated a healthy voice in both binary and multi-class classifications. In multi-class classification, 77 % of cases with short activation time were healthy voices, which is slightly lower than the 100 % healthy voices in binary classification, but still a very high percentage. Discussion There are various methods for analyzing sound. Recently, analysis methods using AI, including machine learning and deep learning, have been advancing [30,31]. For this, the digital conversion of sound is essential. In sound research, particularly in speech recognition, MFCCs has been widely used. As a result, compared to other filters that use a linear frequency scale, the MFCC uses logarithmic scale and includes a relatively large number of frequency components in low-frequency bands and a smaller number of frequency components in high-frequency bands. The logarithmic scale, or mel scale, is suitable for human pitch interpretation [10,12,17]. Similarly, OFSE use octave filter bank, meaning the upper frequency limit is twice the lower frequency limit. OFSE, especially OFSE with 1/3 octave band filter, is also a logarithmic frequency resolution. Furthermore, this provides finer resolution at lower frequencies, especially under 200 Hz. According to the source-filter theory [29], the larynx is the organ that creates the sound source. Voice is produced by the contact of the vocal folds, and the two vocal folds produce the fundamental frequency through two major factors: One is the differences in muscle and mucosal contraction, and the other is the Bernoulli effect caused by the speed of air passing from the lungs through the trachea [32]. This fundamental frequency is then resonated in the vocal tract, specifically in the pharynx and oral cavity. Therefore, phonetic characteristics of laryngeal diseases can affect the fundamental frequency and acoustic parameters such as jitter and shimmer, which are variations in the period and amplitude of the fundamental frequency more than several factors in the formants [33]. In this study, laryngeal diseases are categorized into three groups. One is vocal fold paralysis, representing weak or absent vocal fold contacts. The other two are benign mucosal disease and laryngeal cancer, including glottic cancer and transglottic cancer. These represent stronger vocal fold contacts and have pathologic differences that change clinical treatment. For a healthy male, the fundamental frequency is typically between 120 and 150 Hz. Even with changes due to laryngeal diseases, the fundamental frequency does not exceed 200 Hz [34]. For this reason, it is crucial to closely examine under 200 Hz initially. A formant is a concentration of acoustic energy around a particular frequency in the speech wave, determined by the width of the pharynx and the length of the oral cavity according to vowel pronunciation. In healthy individuals, vocal fold vibration is symmetrically aligned on both sides, producing a clear fundamental frequency for a prolonged period. Formants also consistently appear within specific narrow ranges while producing particular vowels However, if there is a problem in vocal folds vibration, significant changes in the fundamental frequency occur and lead to irregular resonance over a broader range of frequencies compared to a healthy voice. Sometimes, this can manifest as increased noise in regions over the formant areas. In summary, using a filter that analyzes more accurately in frequencies up to 200 Hz, which includes the fundamental frequency, and covers the formant area and over-formant area up to 20,000 Hz, allows for a more accurate distinction of voice changes due to pathological changes in the larynx. OFSE with a 1/3 octave band filter differentiates the region below 200 Hz using 12 index filters, whereas MFCC uses only 2 index filters. The first and second formant, 700–1500 Hz, is divided into 5 index filters by both MFCC and OFSE. This indicates there is not significant difference in the formant area. From this perspective, OFSE with 1/3 octave band filter can be considered a superior filter compared to MFCC. Comparing OFSE with 1/3 octave band filter and MFCC, binary classification showed similarly high results for both. In classification between healthy and laryngeal caner group, MFCC showed a very high accuracy of 97 %, but OFSE demonstrated an accuracy of 100 %. Although OFSE with 1/3 octave band filter is slightly higher than MFCC, the difference is not statistically significant. However, when comparing healthy with vocal fold paralysis and healthy with benign mucosal diseases conditions, OFSE showed a significantly higher with an accuracy of 98–100 %. In contrast, MFCC distinguished healthy voices and laryngeal cancer voices with an accuracy of over 97 %, but its accuracy dropped to 88–90 % for vocal fold paralysis and benign mucosal diseases. According to the MFCC results, laryngeal cancer is the easiest to distinguish from healthy among benign mucosal disease, vocal fold paralysis, and laryngeal cancer voices. Additionally, the difference between results of OFSE and MFCC in distinguishing benign mucosal disease and vocal fold paralysis implies that these two conditions require the use of both the formant area and the fundamental frequency area for accurate differentiation more than the laryngeal cancer condition. In multi-class classification, OFSE with 1/3 octave band filter showed better results than MFCC. OFSE with a 1/3 octave band filter showed an accuracy of over 92 %, whereas MFCC demonstrated an accuracy of 70 %. These results, similar to those in binary classification, can be attributed to the analysis utilizing not only the formant area but also the fundamental frequency area. Hung et al. [13] and Hu et al. [35] have reported an accuracy rate of 77.5 % and 66.9 %, respectively, categorizing voices into five classes, including neoplasm. Studies attempting to distinguish laryngeal cancer or neoplasm using MFCC have shown accuracy rates ranging from 66 % to 77 %. Although there are not many studies, the fact that the accuracy of previous models did not exceed 80 %, whereas the new model shows an accuracy of over 90 %, is a very significant result. In this study, we analyzed the scope of OFSE using Grad-CAM. When distinguishing between laryngeal cancer and healthy voice, including overlaps, 48 % of the over-formant area was utilized. The formant area was utilized about 42 %, and the under-formant area, which includes the fundamental frequency, was used 34 %. However, the characteristics of laryngeal cancer voice showed that 74 % was in the over-formant area, and 26 % was in the under-formant area. This indicates that OFSE judged based on the increased noise in the over-formant area and changes in the fundamental frequency range in laryngeal cancer voices compared to healthy ones. Although the formant area was used about 42 %, this was entirely used for distinguishing healthy voices, a figure arising due to the larger number of normal samples. In multi-class classification, the utilization of the over-formant area decreased to 28 %, the formant area increased to 57 %, and the under-formant area was utilized about 24 %. The increase in the formant area utilization is likely because it is necessary to distinguish among various laryngeal disease conditions, not just between healthy and abnormal groups. Laryngeal disease voices make some noises in over-formant area. However, because noise typically lacks a discernible pattern, it is challenging to classify. In binary classification, noise can simply be identified as laryngeal disease voice, but in multi-classification, it is necessary to examine both the formant and fundamental frequency areas to distinguish among laryngeal disease voices. Using Grad-CAM, we have found several interesting results. First, the under-formant area, including the fundamental frequency, was significantly utilized. The under-formant area, including the fundamental frequency, was utilized 34 % in binary classification and 24 % in multi-classification. These are by no means insignificant figures, indicating that this area plays a crucial role in distinguishing laryngeal diseases. The difference in accuracy between OFSE and MFCC can be confirmed in this area. Second, despite training with short period of 0.5 s, there are some cases that only specific short period was important for disease differentiation. Healthy voices can be differentiated by only for relatively short activated period. According to Fig. 17, in clusters where only short period was activated, such as early narrow formant area enhancement, late formant & over-formant area enhancement, and single under-formant area enhancement, 375 out of 755 healthy voices, corresponding to 49.7 %, were identified. The position of the activated area—whether at the beginning or end of the voice—did not make a difference. This is because, when subdividing a 4-s/a/sound, the strong characteristics of a healthy voice can be captured in either the beginning or the end. Third, specific narrow frequency areas were often important. Particularly vocal fold paralysis was often diagnosed based solely on narrow areas, regardless of whether they were in the formant, over-formant, or under-formant areas. These differences are also perceptible to the expert human ear and are clinically expressed using scores like the GRBAS score [34]. Forth, there was a cluster without any enhancement, indicating a lack of special features. Among the voices in this cluster, 86 % were healthy voices and the others were benign mucosal disease, with no cases of laryngeal cancer or vocal fold paralysis. Healthy voices are relatively consistent, producing normal sound from beginning to end without changes, indicating that Grad-CAM can recognize this. Additionally, this suggests that benign mucosal diseases, though rare, can produce sound similar to that of a healthy voice. In contrast, even when broken down into very short segments, laryngeal cancer does not produce a normal waveform. One of the most challenging tasks is distinguishing between benign mucosal diseases and laryngeal cancer. The significance of Grad-CAM's analysis is that it offers the potential to present the principle of how AI distinguishes between benign mucosal diseases and laryngeal cancer. Limitation of the study There are several limitations. One is that the study was conducted at a single center, with no external validation. The next study needs multi-center data and external validation. Secondly, there is a difference in sample sizes due to the prevalence of diseases, with a smaller number of cases for laryngeal cancer. For the same reason, the study was conducted only on men, excluding women. Additionally, while we interpreted CNN predictions using CAM to understand the performance improvement with OFSE, further development of models that can more seamlessly integrate with clinical and phonetic analysis is needed. Future work Future research should explore the application of OFSE in combination with other advanced deep learning models to further improve diagnostic accuracy and reliability. For example, we will explore the application of OFSE in combination with transformer-based or physics-informed neural networks. Also, integrating OFSE with explainable AI techniques could enhance the interpretability of the model's decision-making process, providing clinicians with clearer insights into the specific features and frequency bands associated with laryngeal diseases. Such future works based on this paper would not only boost performance but also support more transparent and clinically meaningful AI applications in voice-based diagnostics. External validation with diverse patient populations is necessary to ensure broader applicability. Additionally, integrating real-time voice analysis tools for early screening and monitoring of laryngeal conditions in clinical practice could be an important step toward improving patient outcomes. Conclusion This study demonstrated that OFSE with 1/3 octave band filters provides superior accuracy compared to MFCC in classifying laryngeal diseases, including laryngeal cancer, benign mucosal disease, and vocal fold paralysis, using voice data. OFSE's ability to capture subtle changes in the fundamental and under-formant frequency areas significantly improves diagnostic performance, especially in multi-class classifications. These findings suggest that OFSE with 1/3 octave band filters can replace MFCC for use in AI-based diagnosis, enabling more accurate, non-invasive early detection of laryngeal diseases. CRediT authorship contribution statement Jaemin Song: Writing – original draft, Methodology, Investigation, Formal analysis, Data curation. Hyunbum Kim: Writing – original draft, Funding acquisition, Formal analysis, Data curation, Conceptualization. Yong Oh Lee: Writing – review & editing, Supervision, Project administration, Methodology, Funding acquisition, Conceptualization. Resource availability Lead contact Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Yong Oh Lee (Email: yongoh.lee@hongik.ac.kr). Materials availability This study did not generate new unique reagents. Data and code availability Voice data reported in this paper will be shared by the lead contact upon request. Information about code will be prepared at github and is publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Funding This work was supported by Hongik University New Faculty Research Fund. This study was supported by a grant of the E.N.T. Fund of the Catholic University of Korea made in the program year of 2024. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements The present study protocol was reviewed and approved by the Institutional Review Board (IRB) and Ethics Committee of The Catholic University of Korea in accordance with the Declaration of Helsinki. (XC23RIDI0042). References Associated Data Data Availability Statement Voice data reported in this paper will be shared by the lead contact upon request. Information about code will be prepared at github and is publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.