AcousticSceneBD: A real-world environmental audio dataset for acoustic scene classification in Bangladesh Authors: Tasnova Haque Mazumder, Akash Ahmed, Dilruba Akter, Amin Ocin, Raihan Ul Islam, Mohammad Rifat Ahmmad Rashid, Shamim H Ripon, Ahmed Wasif Reza Published: 2026 ABSTRACT AbstractEnvironmental sound classification (ESC) has been used to implement various applications such as smart city, environmental monitoring systems, context-aware systems, and assistive technologies. The existing benchmark datasets, however, mainly focus on the acoustic environment of the west and limited acoustic scenes are available in Bangladesh. In this paper, we introduce a real-world environmental audio dataset named AcousticSceneBD, with a total of 5535 recordings across 7 classes of acoustic scenes: Bus, Metro, Metro Station, Park, Restaurant, Shopping Mall, and University. All recordings were conducted using the standard operating conditions of the smartphone's microphone and with all files standardized following the same procedure:10-second, mono, 16-bit WAV audio recorded at 44.1/48 kHz, with a total duration of 15.38 hours and a dataset size of 4.87 GB.To guarantee consistency and reproducibility, the dataset has been created using a structured approach for data collection and processing consisting of eight stages. Four baseline models, namely Random Forest, YAMNet, Wav2Vec2, and PANNs CNN14 were tested to validate the data, with respective test accuracy of 58.81%, 82.99%, 47.11%, and 78.15%, where Random forest is trained on extracted MFCC features from the raw audio The results indicate that AcousticSceneBD is well annotated, acoustically discriminative and appropriate for the development and evaluation of traditional, deep learning, transfer learning, and self-supervised environmental sound classification methods.Keywords: Acoustic dataset, Environmental sound classification, Audio signal processing, Deep learning, Urban soundscape, Transfer learning, Artificial intelligence, Bangladesh Specifications Table FieldDescriptionSubjectComputer SciencesSpecific subject areaAudio Signal Processing, Environmental Acoustics Scene, Sound Classification, Artificial IntelligenceType of dataAudio (.wav, mono, 16-bit, 44.1/48 kHz); Comma-separated metadata (.csv);Data collection5535 audio clips of 10-second duration recorded across seven environmental classes (Bus, Metro, Metro Station, Park, Restaurant, Shopping Mall, University). Collected over multiple field-recording sessions in Bangladesh using commercial smartphone microphones (Samsung and Realme handsets) placed at approximately human ear height (∼1.5 m), with peak meters kept between -12 and -6 dB to prevent clipping. Raw recordings were segmented into non-overlapping 10-second chunks, peak-normalized to -3 dBFS, and stored as 16-bit mono WAV files.Data source locationUrban and institutional environments in Bangladesh (city bus routes, metro lines and stations, restaurants, shopping malls, urban parks, and university campuses, primarily in Dhaka). Affiliation: East West University, Aftabnagar, Dhaka 1212, Bangladesh.Data accessibilityRepository name: Mendeley DataDOI: 10.17632/bynmzy4dp6.1Direct URL: https://data.mendeley.com/datasets/bynmzy4dp6/1Related research articleNone. Value of the Data •AcousticSceneBD aims to fill a geographical gap in the realm of environmental audio, by offering high-quality recordings from a wide range of Bangladeshi urban, transportation, commercial, institutional and recreational soundscapes, which are underrepresented in widely used benchmark datasets. AcousticSceneBD aims to fill a geographical gap in the realm of environmental audio, by offering high-quality recordings from a wide range of Bangladeshi urban, transportation, commercial, institutional and recreational soundscapes, which are underrepresented in widely used benchmark datasets. •To the best of our knowledge, AcousticSceneBD is one of the first publicly available environmental audio datasets collected for Bangladesh that emphasizes real-world acoustic scene classification from a variety of real-world environments. To the best of our knowledge, AcousticSceneBD is one of the first publicly available environmental audio datasets collected for Bangladesh that emphasizes real-world acoustic scene classification from a variety of real-world environments. •All 5535 clips are of the same length (10 seconds), allowing them to be seamlessly incorporated into common machine learning workflows without extra segmentation or preprocessing. All 5535 clips are of the same length (10 seconds), allowing them to be seamlessly incorporated into common machine learning workflows without extra segmentation or preprocessing. •The consistent recording and processing specifications allow the extraction of features from spectrograms and Mel-spectrograms, MFCCs, and recent self-supervised audio representation in a reproducible manner. The consistent recording and processing specifications allow the extraction of features from spectrograms and Mel-spectrograms, MFCCs, and recent self-supervised audio representation in a reproducible manner. •The data set is validated with several baselines (Random Forest, YAMNet, Wav2Vec2, and PANNs CNN14) to further establish its suitability for research on environmental sound classification and to give benchmarked results for future studies. The data set is validated with several baselines (Random Forest, YAMNet, Wav2Vec2, and PANNs CNN14) to further establish its suitability for research on environmental sound classification and to give benchmarked results for future studies. •The presented dataset can be used in future research to develop adaptive hearing aids, assistive listening devices, environmental awareness systems, smart-city monitoring systems, and context-aware audio applications, offering representative recordings of typical acoustical environments. The presented dataset can be used in future research to develop adaptive hearing aids, assistive listening devices, environmental awareness systems, smart-city monitoring systems, and context-aware audio applications, offering representative recordings of typical acoustical environments. •AcousticSceneBD can be extended with more locations, environmental categories, seasons and recording situations which can be used as a basis for a complete acoustic benchmark of acoustic scenes and acoustic analyses of soundscape in Bangladesh. AcousticSceneBD can be extended with more locations, environmental categories, seasons and recording situations which can be used as a basis for a complete acoustic benchmark of acoustic scenes and acoustic analyses of soundscape in Bangladesh. Background The field of environmental sound classification (ESC) has become an essential research area in the field of audio signal processing and artificial intelligence, finding applications in smart cities, environmental monitoring, surveillance systems, context-aware computing, and assistive technologies. Intelligent systems can better understand and adapt to the environment when automatically recognizing acoustic environments from short audio recordings. There are a number of datasets of environmental audio that have been introduced for the benefit of ESC research (Table 1), but they share some limitations. The majority of the publicly available datasets are dominated by recordings from North American and European areas, with acoustic environments for Bangladesh being under-represented. Consequently, sound attributes of local transportation systems, commercial facilities, educational facilities, and public recreational areas are rarely provided in benchmarks. In addition, a large proportion of the datasets are based on crowdsourced recordings collected from websites, and these may result in variations in quality of the recording, reliability of the annotations, recording equipment, and environment in which they are recorded. Moreover, there are some datasets with variable-length audio files which need further segmentation, padding, or pre-processing before they can be successfully used in state-of-the-art machine-learning applications, which may have an impact on reproducibility or benchmarking consistency. Table 1Comparison with existing acoustic environment / sound-classification datasets.Table 1 dummy alt textDatasetClassesLocation / SourceClipsDuration (h)ESC-10 [1].10Global (Freesound)4000.55UrbanSound8K [2].10USA (urban)8,7328.75TUT Acoustic Scenes [3].10Finland10,52029.20GESMA [4].3 (hier.)Ghana (urban/inst.)22,19362.24AcousticSceneBD (This Work) [5].7Bangladesh (urban/inst.)553515.38 Comparison with existing acoustic environment / sound-classification datasets. To overcome these shortcomings, we have created AcousticSceneBD, a real-world environmental sounds dataset gathered from a variety of daily life environments in Bangladesh. We have 5535 standardized and cut 10-second audio clips from seven acoustic scenes: Bus, Metro, Metro Station, Restaurant, Shopping Mall, University, and Park. All recordings were taken under natural operating conditions in the real environment using a smartphone microphone, which was able to record the real environment's sounds and acoustic variability. AcousticSceneBD is a helpful resource for environmental sound classification, audio representation learning, transfer learning and self-supervised learning research. In addition, the dataset can provide future applications in smart city, environmental surveillance, context-aware audio systems and assistive listening technologies. Data Description The AcousticSceneBD dataset consists of 5,535 audio samples arranged in the following folders: Bus, Metro, Metro Station, Restaurant, Shopping Mall, University, and Park each containing audio samples by class. All recordings are recorded for 10.0 seconds so that the recordings are all the same length and do not need to be segmented or padded into the machine learning pipelines. All recordings were made in real-world situations with the built-in mic of the Samsung and Realme phones. To account for natural variability in the device and to represent realistic recording characteristics, multiple smartphone models were used. All audio files recorded are 16-bit, mono, lossless PCM WAVs at either 44.1 kHz or 48 kHz sample rate due to different device recording configurationsThis may lead to some differences in the spectral features as the baseline inputs were resampled to match the requirements of each model. The released files and metadata contain the original sampling rates to ensure the data can be reproduced.The average file size is about 0.88 MB, the size of all the files on disk is about 4,866.39 MB (4.87 GB), and the total duration is about 922.50 minutes (15.38 hours). The dataset was limited to 7 classes to emphasize more diverse and representative real-world recordings of common soundscapes in Bangladesh, rather than aiming to match the number of classes found in existing benchmark datasets. Table 2 lists the number of clips, percentage of the dataset, total duration and on-disk size for each class. The distribution of the clips in the class is displayed in Fig. 1, and the detailed technical parameters of AcousticSceneBD are summarized in Table 3. Table 2Number of clips, percentage of dataset, total duration, and on-disk size per class.Table 2 dummy alt textClassClips% of datasetDuration (min)Size (MB)Bus1,43225.87%238.671,204.58Metro57610.41%96.00527.37Metro Station62911.36%104.83575.90Park4968.96%82.67454.12Restaurant98817.85%164.67886.06Shopping Mall69012.47%115.00609.35University72413.08%120.67609.02Total5,535100.00%922.50 (15.38 h)4,866.39 Number of clips, percentage of dataset, total duration, and on-disk size per class. Class distribution: number of clips per class (left) and total duration per class in minutes (right). Table 3Technical specifications of AcousticSceneBD.Table 3 dummy alt textParameterValueTotal audio files5535Number of classes7Clip duration10.0 seconds (fixed-length)Total duration15.38 hours (922.50 minutes)Audio formatWAV (lossless PCM)Sample rate44.1 kHzBit depth16-bitChannelsMonoAverage file size∼0.88 MBTotal dataset size4,866.39 MB (∼4.87 GB)Label typeSingle-label (one class per file)Recording devicesSamsung & Realme smartphones Technical specifications of AcousticSceneBD. Folder structure The data is arranged as a flat tree with one subfolder for each class as illustrated in Fig. 2. The name of each clip is -_chunk_.wav, where records the session this clip was recorded in and indicates the 10-second clip within the session. This is accompanied by a per file metadata file (metadata.csv) which contains information about the file such as class, duration, sample rate, channels, frame count, and on-disk size: Folder structure visualization. Acoustic classes Typical sources, frequency range, temporal pattern, spectral features and a distinguishing feature are discussed below for each of the scene classes. “Fig. 3 shows representative waveforms for each class, and Fig. 4 shows the class-level spectral feature distributions. Representative 10-second time-do main waveforms, one example per class. Distributions of spectral centroid, zero-crossing rate, spectral rolloff, and RMS energy across the seven classes. Bus Diesel and CNG city bus engine noise, gear changes, door opening/closing, brake noises, tire-road noise and short passages of passenger voices recorded in moving buses as well as at bus stops. •Frequency range: 100Hz to 8kHz with a predominance of low frequencies below 1kHz. Frequency range: 100Hz to 8kHz with a predominance of low frequencies below 1kHz. •Temporal pattern: Continuous broadband engine drone, periodic cycling and impulsive door/gear events. Temporal pattern: Continuous broadband engine drone, periodic cycling and impulsive door/gear events. •Spectral characteristics: Well-defined harmonics at the engine's RPM; strong low-frequency rumble. Spectral characteristics: Well-defined harmonics at the engine's RPM; strong low-frequency rumble. •Distinguishing feature: It has the lowest spectral centroid of any class and a characteristic diesel-engine signature. Distinguishing feature: It has the lowest spectral centroid of any class and a characteristic diesel-engine signature. Samples: 1432 clips, 238.67 min (3.97h). Metro On-board metro train sounds recorded during the acceleration, cruising and braking phases. •Frequency range: 200 Hz to 5 kHz. Frequency range: 200 Hz to 5 kHz. •Temporal pattern: Continuous cruise with fast amplitude modulation during acceleration / braking; rail-joint impulses. Temporal pattern: Continuous cruise with fast amplitude modulation during acceleration / braking; rail-joint impulses. •Spectral characteristics: harmonic content with metallic-rail resonances and a whine from the traction motors. Spectral characteristics: harmonic content with metallic-rail resonances and a whine from the traction motors. •Distinguishing feature: Is higher pitched than bus, but has high frequency clicks periodically. Distinguishing feature: Is higher pitched than bus, but has high frequency clicks periodically. Samples: 576 clips, 96.00 min (1.60 h). Metro station Ambient platform background noise such as PA announcements, train arrivals/ departures, crowd noise and footstep noise. •Frequency range: 300 Hz to 6 kHz. Frequency range: 300 Hz to 6 kHz. •Temporal pattern: Reverberant, intermittent loud events on continuous background of crowd noise. Temporal pattern: Reverberant, intermittent loud events on continuous background of crowd noise. •Speech characteristics: Broad-spectrum noise, speech formants, large tail of the reverberation. Speech characteristics: Broad-spectrum noise, speech formants, large tail of the reverberation. •Distinguishing feature: High zero crossing rate and good reverberant signature, PA tones audible. Distinguishing feature: High zero crossing rate and good reverberant signature, PA tones audible. Samples: 629 clips, 104.83 min (1.75 h). Park Natural soundscape outdoors with bird calls, wind through foliage and occasional distant traffic noise. •Birdsong: 2 kHz – 8 kHz; Wind rumble: 50 Hz – 200 Hz. Birdsong: 2 kHz – 8 kHz; Wind rumble: 50 Hz – 200 Hz. •Temporal pattern: Intermittent tonal bird calls over a low energy background. Temporal pattern: Intermittent tonal bird calls over a low energy background. •The features of the spectrally analysed events are: Tonal, narrow-band bird events, low-frequency wind energy. The features of the spectrally analysed events are: Tonal, narrow-band bird events, low-frequency wind energy. •Distinguishing feature: Least overall RMS energy in the dataset; sparse, natural soundscape. Distinguishing feature: Least overall RMS energy in the dataset; sparse, natural soundscape. Samples: 496 clips, 82.67 min (1.38 h). Restaurant Crowd noise, cutlery noises, kitchen noises, music and service announcements in an indoor restaurant. •Frequency range: 500 Hz to 8 kHz. Frequency range: 500 Hz to 8 kHz. •Temporal pattern: Continuous crowd noise with sharp impulsive cutlery/dish sounds. Temporal pattern: Continuous crowd noise with sharp impulsive cutlery/dish sounds. •Spectral characteristics: Glass and metal has transient high frequency peaks while dense speech formants. Spectral characteristics: Glass and metal has transient high frequency peaks while dense speech formants. •Distinguishing Feature: Speech Band Energy is high and characteristic of clatter transients. Distinguishing Feature: Speech Band Energy is high and characteristic of clatter transients. Samples: 988 clips, 164.67 min (2.74 h). Shopping mall Ambient background noise in large indoor commercial areas such as crowd murmur, retail announcement, escalator/HVAC noise, and background music. •Frequency range: 300 Hz to 5 kHz. Frequency range: 300 Hz to 5 kHz. •Temporal pattern: Flow of crowd talking continuously with some occasional PA announcements. Temporal pattern: Flow of crowd talking continuously with some occasional PA announcements. •Reverberant indoor characteristics, distributed speech energy, low-frequency HVAC hum. Reverberant indoor characteristics, distributed speech energy, low-frequency HVAC hum. •Lower transient density and longer reverberation tail than Metro Station is the distinguishing feature. Lower transient density and longer reverberation tail than Metro Station is the distinguishing feature. Samples: 690 clips, 115.00 min (1.92 h). University In the interior and exterior of the campus, such as corridors, between lectures, library lobby, and student gathering space. •Frequency range: 200 Hz to 4 kHz. Frequency range: 200 Hz to 4 kHz. •Temporal pattern: Interrupted activity at the time of the transition into and out of class; calmer times elsewhere. Temporal pattern: Interrupted activity at the time of the transition into and out of class; calmer times elsewhere. •Spectral features: Speech dominated with footstep and door transients. Spectral features: Speech dominated with footstep and door transients. •The distinguishing feature is smaller crowd densities and less reverberation than Mall/Restaurant. The distinguishing feature is smaller crowd densities and less reverberation than Mall/Restaurant. Samples: 724 clips, 120.67 min (2.01 h). Metadata description The metadata file contains detailed information for each audio clip, such as its name, class label, file size, number of channels, sampling rate, duration and number of frames. These metadata enable organization, verification, and reproducible pre-processing of datasets.The metadata fields and their descriptions are summarized in Table 4, while Table 5 presents representative example rows from the metadata file. Table 4Description of metadata columns.Table 4 dummy alt textColumnDescriptionfilenameName of the audio file.relative_pathRelative path of the audio file within the dataset directory.class_labelAcoustic-scene category assigned to the audio clip.file_size_bytesSize of the audio file in bytes.file_size_mbSize of the audio file in megabytes.channelsNumber of audio channels; a value of 1 indicates mono audio.sample_rateSampling rate of the audio clip in Hz.duration_secondsDuration of the audio clip in seconds.num_framesTotal number of audio frames/samples in the clip. Description of metadata columns. Table 5Example rows from the metadata file.Table 5 dummy alt textfilenamerelative_pathclass_labelfile_size_mbchannelssample_rateduration_secondsnum_framesBus-1_chunk_1.wavBus/Bus-1_chunk_1.wavBus0.841214410010441000Metro-1_chunk_1.wavMetro/Metro-1_chunk_1.wavMetro0.915614800010480000Restaurant-2_chunk_1.wavRestaurant/Restaurant-2_chunk_1.wavRestaurant0.915614800010480000 Example rows from the metadata file. Experimental Design, Materials and Methods Data collection pipeline The data collection process was broken down into an eight stage process: (i) site selection across the 7 target classes, (ii) requesting access and calibration of the smartphones, (iii) field recording sessions, (iv) raw data collection of audio, (v) quality filtering, (vi) class labelling and class verification, (vii) 10-second segmentation, and (viii) format standardisation prior to final release summarized in Fig. 5. End-to-end data-collection and processing pipeline for the environmental sound dataset. Site selection Acoustic variability within classes and coverage of the seven target classes was maximized in the selection of sites. To ensure that the data reflected real-world variation in devices, time of day and crowd density (avoiding just the acoustic response of a single location), multiple recording locations were chosen for each class (e.g., multiple bus routes and bus stops for the Bus class, multiple metro stations and trains for Metro class and Metro Station class, multiple parks, restaurants, malls, and university buildings for other classes). Recording equipment Equipment specifications are summarized in Table 6. Table 6Recording equipment.Table 6 dummy alt textEquipmentTypeSpecificationsField recorderSmartphone (Samsung / Realme)Built-in microphone, 16-bit PCM, 44.1/48 kHzRecording appStandard smartphone voice recorderLossless WAV exportWindshieldFoam windshield (where required)Outdoor wind-noise suppressionStorageInternal device storage / SDHC32 GB+ per sessionWorkstationPython / librosa [6]/ DAW for post-processingTrimming, normalization, format standardization Recording equipment. Recording protocol •Session time: 30-60 minutes at each location per session. Session time: 30-60 minutes at each location per session. •Microphone position: Human-ear level (about 1.5 m above ground) and used without any hand-held support to reduce handling noise. Microphone position: Human-ear level (about 1.5 m above ground) and used without any hand-held support to reduce handling noise. •Gain settings: default to the app and manually verified between -12 dB to -6 dB to make sure there is no clipping. Gain settings: default to the app and manually verified between -12 dB to -6 dB to make sure there is no clipping. •Peak Activity Hours: Hours in which there is high activity for each environment (Commute hours for Bus/Metro, Meal hours for Restaurant, Business hours for Mall, Class transitions for University). Peak Activity Hours: Hours in which there is high activity for each environment (Commute hours for Bus/Metro, Meal hours for Restaurant, Business hours for Mall, Class transitions for University). •Environmental conditions: recorded in normal conditions; conditions of abnormal weather, equipment fault, unusual incidents were discarded. Environmental conditions: recorded in normal conditions; conditions of abnormal weather, equipment fault, unusual incidents were discarded. •Background noise: not artificially removed - ambient noise as part of the authentic soundscape. Background noise: not artificially removed - ambient noise as part of the authentic soundscape. Audio processing 1.Segmentation: The first step was to divide each recording of the session into non-overlapping 10-second segments. Segmentation: The first step was to divide each recording of the session into non-overlapping 10-second segments. 2.Normalized: peak amplitude scaled to -3 dBFS to avoid clipping, while maintaining the class-to-class dynamic range. Normalized: peak amplitude scaled to -3 dBFS to avoid clipping, while maintaining the class-to-class dynamic range. 3.Format standardization: all clips were converted to 16-bit mono WAV, and at the original sample rate (44.1 or 48 kHz). Format standardization: all clips were converted to 16-bit mono WAV, and at the original sample rate (44.1 or 48 kHz). 4.Quality filtering: clips that contained clipping, drop-outs, ambiguous content, or contained accidental personally identifiable information were eliminated. Quality filtering: clips that contained clipping, drop-outs, ambiguous content, or contained accidental personally identifiable information were eliminated. 5.Metadata generation: for each clip recorded, we stored: on-disk size, channel count, sample rate, duration, the number of frames, and the class label, in a single CSV metadata file. Metadata generation: for each clip recorded, we stored: on-disk size, channel count, sample rate, duration, the number of frames, and the class label, in a single CSV metadata file. Annotation workflow and inter-annotator agreement The annotation process started while collecting data, and the audio recordings were taken directly from 7 real-world acoustic environments: Bus, Metro, Metro Station, Park, Restaurant, Shopping Mall, and University. Each of the recorded audio files was then inspected, by listening to the acoustics of the recording, and labeled with the relevant scene label, depending on the environment in which the recording was made and on the predominant sounds in the audio file. A random sub-sample of 350 audio clips (50 audio clips from each of the seven classes) was selected for the purpose of assessing the reliability of the annotations. All selected clips were played to two annotators who independently labelled them according to the pre-defined scene categories. To not influence the other annotator, independent annotation was used. There were 350 clips with a raw inter-annotator agreement of 64.29%, between the two annotators. In order to correct for chance agreement, Cohen's kappa (κ) was also computed. The inter-annotator agreement was found at a moderate level of 0.5829 with the obtained value. The manual annotations were also matched with the existing set of data labels. The percentage of agreement between Annotator 1 and the rest of the community was 71.14%, and between Annotator 2 and the rest of the community was 74.57%. The majority of the disagreements were found between acoustically similar environments and when sounds overlap, such as voices in conversations, crowd noise, music or announcements. Validation with baseline models Four different baseline models were trained and tested to ensure that the data set is well labeled and learnable. The four baselines are (i) classical feature engineering with a Random Forest (RF), (ii) supervised transfer learning from a model pre-trained on a large-scale general audio corpus (YAMNet), (iii) self-supervised speech embeddings (Wav2Vec2), and (iv) end-to-end audio-domain CNN pre-training (PANNs CNN14). To prevent data leakage caused by acoustic overlap across continuous recording sessions, the dataset was partitioned strictly at the recording-session level rather than splitting individual clips. Recording sessions within each class were stratified and assigned to train, validation, or test partitions targeting a 70/10/20 ratio, after which all clips belonging to each session were merged to form the final datasets. Holding out complete, un-overlapped recording sessions yielded 3,032 clips for training, 1,221 clips for validation, and 1,282 clips for testing (out of 5,535 total clips). This structure ensured that no temporally adjacent or acoustically shared background audio contaminated the evaluation splits.. The following table (Table 7) shows supports provided per class. Table 7Per-class precision (P), recall (R) and F1-score for each baseline model. 'Sup' is the test-set support per class. The last row reports the support-weighted average F1.Table 7 dummy alt textClassSupRF PRF RRF F1YAMNet PYAMNet RYAMNet F1W2V PW2V RW2V F1PANNs PPANNs RPANNs F1Bus3610.9810.8590.9160.9940.9750.9850.710.7810.7440.9780.9720.975Metro1500.9740.980.9770.9660.9330.9490.4810.5130.4970.9660.9330.949Metro Station890.4890.9890.6540.8530.9780.9110.2860.7640.4160.5670.9550.711Park1390.8530.5830.6920.8150.8850.8480.2990.1440.1940.8870.6190.729Restaurant2490.0390.0080.0130.6590.9160.7660.4060.2090.2760.6450.8920.749Shopping Mall1090.0220.0730.0340.6880.5870.6340.0910.1190.1030.4960.5050.5University1850.9920.6380.7760.7690.3780.5070.6170.4970.5510.8290.3410.483Macro avg F11282——0.58——0.8——0.397——0.728Weighted avg F11282——0.61——0.82——0.459——0.772 Per-class precision (P), recall (R) and F1-score for each baseline model. 'Sup' is the test-set support per class. The last row reports the support-weighted average F1. Random forest on MFCC features For every clip, we calculated 40 dimensional MFCCs on 25 ms of windows with a 10 ms hop step, then summarized each of the 40 MFCCs coefficients with four statistics (mean, std, max, min) that gave us 160 dimensions, plus delta and delta-delta statistics to give us a final feature vector of 177 dimensions. This feature matrix was used to train a Random Forest classifier with n_estimators = 200, with no depth limit applied, and random_state = 42. Test accuracy = 58.81%. Low-order MFCC means and standard deviations (Fig. 6) dominate the most important features. Top-30 random forest feature importances. Low-order MFCC means and standard deviations dominate. YAMNet + MLP (transfer learning) YAMNet [7]. is a mobile version of a pre-trained model on a large-scale general set of audio data, which generates 1024-dimensional embedding every ∼0.96 s frame. For each clip we resampled audio to 16 kHz mono and mean-pooled YAMNet frame embeddings to obtain a single 1024-dimensional vector. The model was trained using Adam (learning rate 1e-3), categorical cross-entropy, ReduceLROnPlateau and EarlyStopping (patience 10) with batch size 32 for a maximum of 50 epochs. The training curves in Fig. 7 yield an accuracy of 82.99% for the test. YAMNet+MLP training accuracy and loss curves over 50 epochs. Wav2Vec2 + MLP (self-supervised representations) Wav2Vec2 [8]. is a self Supervised Speech Encoder. After resampling each clip to 16 kHz mono, we extracted 768 dimensional embeddings from each clip using the publicly available wav2vec2-base-960h checkpoint, by taking the mean of the embeddings over time. The embeddings were normally scaled and passed into the 3-layer MLP head as was done in YAMNet. The accuracy of the test set = 47.11% (training curves in Fig. 8). The lower accuracy of Wav2Vec2 (47.11%) indicates a possible domain mismatch, because the Wav2Vec2 model was mainly trained on speech representations (LibriSpeech, not on environmental sounds. The results presented here suggest that speech-based models may need much more fine-tuning on domain-specific acoustic scenes to get a good grasp of environmental acoustics, but it is hoped that this hypothesis can be further tested by controlled transfer experiments (such as cross-dataset pretraining). Wav2Vec2+MLP training accuracy and loss curves. PANNs CNN14 + MLP (end-to-end audio deep learning) PANNs CNN14 [9]. is a 14-layer CNN pre-trained on a large-scale general audio corpus, with log-mel spectrogram inputs. We used 64-band log-mel spectrograms, computed during training using SpecAugment [10]. and 2048-dimensional penultimate-layer embeddings, sampled at 32 kHz. The same 3-layer MLP classification head was added. The accuracy of the tests is 78.15% (training curves in Fig. 9). PANNs CNN14+MLP training accuracy and loss curves. Per-class metrics and confusion matrices Table 7 shows per-class precision, recall, and F1-score of the four baseline models evaluated on the AcousticSceneBD test set. All four models' held-out test set confusion matrices are put together in Fig. 10. Test-set confusion matrices for the four baseline models. Aggregate model comparison Summary of the aggregate test accuracies are shown in Fig. 11 and Table 8. Among the four models, the highest test accuracy achieved was 82.99%. The under-performing Wav2Vec2 baseline is a useful negative control: it shows that the data is non-trivial for representation learners pre-trained on data out of the audio-scene domain, and provides margin for future research in self-supervised audio pre-training specifically geared towards environmental sound. Aggregate test-set accuracy of the four baseline models. Table 8Summary of baseline model results.Table 8 dummy alt textModelTest AccuracyWeighted F1Feature paradigmRandom Forest (MFCC, 200 trees)58.81%0.580Classical featuresYAMNet + MLP82.99%0.800General-audio CNN transfer learningPANNs CNN14 + MLP78.15%0.727Audio-domain CNN pre-trainingWav2Vec2 + MLP47.11%0.397Self-supervised speech encoder Summary of baseline model results. Limitations ●The dataset consists of seven acoustic-scene classes, and not all the different acoustic environments present in Bangladesh are covered. The dataset consists of seven acoustic-scene classes, and not all the different acoustic environments present in Bangladesh are covered. ●Data is geographically and temporally restricted to the locations and time periods where it was gathered and recorded as detailed in the Methods - generalization to other areas, seasons, devices, and acoustic conditions has not been concluded. Data is geographically and temporally restricted to the locations and time periods where it was gathered and recorded as detailed in the Methods - generalization to other areas, seasons, devices, and acoustic conditions has not been concluded. ●Class frequencies are not equal and therefore macro-averaged, balanced, weighted and per-class metrics are presented. Class frequencies are not equal and therefore macro-averaged, balanced, weighted and per-class metrics are presented. ●More than one original sampling rate are in the source files, and each model input is resampled using express model-specific resampling. More than one original sampling rate are in the source files, and each model input is resampled using express model-specific resampling. ●The benchmark models are not an architecture or hyperparameter search, but validation baselines. The benchmark models are not an architecture or hyperparameter search, but validation baselines. ●Causal explanation of cross-domain behavior is not part of this dataset-validation benchmark, due to the fact that the performance of Wav2Vec2 alone is not a sufficient causal explanation. Causal explanation of cross-domain behavior is not part of this dataset-validation benchmark, due to the fact that the performance of Wav2Vec2 alone is not a sufficient causal explanation. ●These recordings have both 44.1 kHz and 48 kHz sampling rates, and could give slightly different spectral representations if features are extracted directly at the sampling rate of the source recordings. These recordings have both 44.1 kHz and 48 kHz sampling rates, and could give slightly different spectral representations if features are extracted directly at the sampling rate of the source recordings. Ethics Statement The authors have adhered to the ethical standards for the publication of data in Data in Brief. There are no experiments on human subjects, animal experiments, or data obtained from social media involved in the current work. All recordings were made in public areas where there was a low expectation of privacy. There was no specific recording, isolation or transcription of identifiable human voice data, such as human speech, or biometric data, such as voice. Since there were no recruitment and interaction with human participants involved in the study, no institutional ethics review was sought. The released data set does not contain any personal information about participants or any intentionally recorded human speech. Sites administration and facility managers have been consulted and permission was granted as necessary. Wildlife recordings were done without baiting, playback or disturbance of wildlife (Park class). The dataset contains no third party copyrighted material and it is available under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Credit Author Statement Tasnova Haque Mazumder: Conceptualization, Methodology, Data curation, Investigation, Validation, Data collection, Writing - original draft, Writing - review & editing. Akash Ahmed: Conceptualization, Methodology, Software, Formal analysis, Supervision, Project administration, Writing - review & editing. Dilruba Akter: Data collection. Amin Ocin: Data collection. Raihan Ul Islam: Conceptualization, Methodology, Supervision, Writing - review & editing. Mohammad Rifat Ahmmad Rashid: Methodology, Supervision, Writing - review & editing. Shamim H Ripon: Supervision, Writing - review & editing. Ahmed Wasif Reza: Supervision, Writing - review & editing.