AudioHate-DB: An audio dataset for English hate speech detection in Bengali accent

Abstract

Hate speech detection has gained significant attention in recent years due to the rapid increase in offensive and harmful content across social media and other digital platforms. Although there are many resources for written hate speech and for accented native English, there are few resources for offensive English speech in South Asian accents, which hinders the development of accent-aware hate speech detection and speech recognition systems. To address this need, the following paper introduces a speech dataset for hate speech detection in spoken English with a Bangladeshi accent, in which slurs, gendered epithets, and general profanity are mixed with otherwise neutral English speech, which is prevalent in Bangladesh and South Asia. This dataset contains 3018 audio samples with a controlled sentence inventory of 1006 sentences (505 hate-class, 501 non-hate-class). Every hateful class sentence has at least one word from a set of 23 hateful terms divided into six semantic categories: racial slurs, gendered/sexual slurs, ability derogatory terms, LGBTQ+ slurs, profanity, and compound/contextual words. Audio samples were collected on mobile telephones using a 44,100 Hz, 16-bit depth format and by multi-regional, multi-gender speakers in natural indoor and outdoor settings and were recorded for two balanced classes. Both raw and noise-cancelled processed versions are offered, where the class name is represented by the directory structure. Baseline experiments using Logistic Regression, Random Forest, Gradient Boosting, and LSTM demonstrate the dataset’s effectiveness, with LSTM achieving the highest accuracy of 83.94%. The dataset can support research on spoken hate speech detection, accent-aware speech recognition, and audio content moderation.

Specifications Table

Subject	Computer Science
Specific subject area	Artificial Intelligence; Audio Processing; Speech Analytics; Machine Learning, Hate Speech Detection; Natural Language Processing.
Type of data	Audio data (wav format)
Data collection	All the data were collected via Android or iOS mobile devices using built-in microphones, resulting in a total of 3018 audio data. All files are stored in uncompressed PCM 16-bit WAV files at a nominal sampling rate of 44,100 Hz (44.1 kHz), which is the same sampling rate as a normal speech-corpus file. The calculated values were intentionally based on regular consumer equipment, thus capturing the variations of the device naturally, the reverberation in the environment, and the ambient noise in the location. These unstructured auditory artifacts are therefore embedded within the corpus, hence accurately reflecting the complex real-world conditions that an operational hate speech detection system must be able to be successful in deployment.
Data source location	The study was conducted in four districts of Bangladesh (Dhaka, Dinajpur, Cumilla, and Khulna), covering the four major zones of the country. These are the regional Bangla accents that add to the regional linguistic and phonetic variations of the speakers.
Data accessibility	Repository name: Mendeley Data Data identification number: [1] https://doi.org/10.17632/hhd57tyy8y.2 Direct URL to data: https://data.mendeley.com/datasets/hhd57tyy8y/2
Related research article	None

1. Value of the Data

• •This dataset provides a foundation for training and evaluating AI models for detecting hate speech in spoken language, featuring a large collection of hate and non-hate utterances in Bengali-accented English, recorded as controlled speech under natural acoustic conditions to ensure accurate classification in authentic real-world scenarios.
• •It will help researchers in the area of speech processing, machine learning, and deep learning to build and improve speech categorization and speech-based content filtering models, including applications in automated content moderation for social media and communication platforms.
• •This dataset can be used for low-resource, culturally sensitive speech research by providing authentic Bengali-accented English speech with common slang as an acoustic alternative to hate speech datasets that are based purely on text, supporting accented-speech research for underrepresented linguistic communities.
• •The inclusion of both raw and noise-reduced audio in this dataset allows researchers to study the effect of pre-processing on model performance and to apply their own custom pipelines.
• •This dataset is a valuable resource for future studies on hate speech detection in audio content and is a stepping stone towards developing speech-based AI tools for real-world communication and social media monitoring.

2. Background

There is a considerable body of research examining hate speech detection with text-based data, especially in high-resource languages, like English. AudioHate-DB consists of audio samples of spoken English hate speech that are spoken with a Bengali accent, while HateXplain is a dataset of English hate speech text documents with explanations labeled by humans [2]. HS-BAN and BanTH offer datasets of Bangla hate speech in text format [3,4], while AudioHate-DB works with spoken English and Bengali-accented speech. While BAAD offers spoken Bangla offensive-language data, AudioHate-DB is specifically focused on detecting hate speech in spoken English with a Bengali accent. AudioHate-DB aims at the classification of hate speech and concentrates on Bengali-accented English speech, while BRADS and BRWDS are mostly used for regional speech research [5,6]. BanglaMUSE offers paired text-audio data, mainly for sentiment analysis [7], while AudioHate-DB has audio data to detect hate speech. Overall, these resources are not similar to AudioHate-DB with respect to language, task, modality, accent, and speaker characteristics. The existing Bangla hate-speech resources are mostly text-based; on the other hand, the existing Bangla speech resources mainly cater to offensive language, sentiment analysis, and regional speech, and not hate-speech detection in English with a Bangla accent. To fill this gap, AudioHate-DB offers controlled recordings of some specific English sentences that are hate speech and non-hate speech sentences spoken by participants as Bengali speakers. The participants had to read the predefined sentences, and the recordings were taken under natural acoustic conditions. The corpus contains both indoor and outdoor recordings and has variation in speaker and acoustic characteristics across various genders. Acoustic features such as MFCC, Chroma-STFT, RMS energy, Spectral centroid, Spectral bandwidth, Spectral roll-off, Zero-crossing rate and Jitter were used for the characterization of the recorded speech signals [8,9]. These features together reflect complementary aspects of the recordings in terms of both spectral, temporal, energy related, and voice quality aspects, thus representing quite comprehensively the underlying acoustic patterns. AudioHate-DB provides a valuable resource for developing and evaluating audio-based hate speech detection systems in Bengali-accented English, addressing the limited availability of resources for this underrepresented linguistic and acoustic setting.

3. Data Description

AudioHate-DB is an audio-based, speaker-diverse, hate speech classification dataset in English language with a Bangladeshi accent. It reflects what is widely seen in Bangladesh and the entire South Asian realm: the intermingling of slurs, gendered epithets, and generic profanity in otherwise neutral English speech in informal conversational or social media settings. Hate speech is described in this study as attacks, demeaning, threatening and encouraging hostility toward a person or group based on an attribute of identity. Just an offensive word was not enough for a Hate label. Rather, the labels were given according to the semantic content and context of the entire sentence. Sentences with offensive words but no hate-related target or intent would not count as Hate. Members of this research group have labelled the dataset into Hate and Non-Hate classes, keeping in mind offensive words and the context of a sentence. Each hate-class sentence contained at least one lexical item from a predetermined hate lexicon consisting of 23 offensive terms organized into six hate semantic categories: racial slurs (4 terms), gendered and sexual epithets (6 terms), ability-derogatory terms (3 terms), LGBTQ+ slurs (2 terms), general profanity (6 terms), and compound (2 terms) or context-dependent derogatory expressions. (Table 1) shows the hate lexicon (HL) classified according to the semantic domain. The dataset is structured with a top-level directory consisting of: Raw/Hate (the raw recordings of hate) and Raw/Non_Hate/ (the raw recordings of non-hate), while /Processed Data/Hate/ contains the noise-reduced versions of /Hate/, and /Processed Data/Non_Hate/ contains the noise-reduced versions of /Non_Hate/. Class membership is then specified in the structure of the folders so that the class labels can be obtained at run-time when loading data, without needing an annotation file. The corpus consists of a controlled sentence inventory with 1006 sentences (505 hate-class, 501 non-hate-class) and a relatively even binary distribution. Along with this, a CSV file has been provided in the directory with Sentence ID, Sentence, Class, Record Type, Recording ID, Duration (seconds), and Division information across three recording cycles (Cycle 1, Cycle 2, and Cycle 3).

Table 1.

Composition of the hate lexicon by semantic category.

Category	Sample Terms	Count
Racial Slurs	nigga, nigger, cracker, honkey	4
Gendered / Sexual	bitch, slut, whore, hoe, cunt, dyke	6
Ability Derogation	moron, retarded, dumb	3
LGBTQ+ Slurs	faggot / fag, tranny	2
General Profanity	fuck, ass, pussy, dick, hell, trash	6
Compound / Contextual	bald bitches, gay (derogatory)	2
Total Hate Lexicon		23

AudioHate-DB consists of 3018 spoken utterances from 100 speakers which includes both male (60%) and female (40%) participants having an average number of recordings of 30, which are recorded from four different regions of Bangladesh, namely Dhaka, Dinajpur, Cumilla and Khulna having 1202, 604, 802 and 410 recordings respectively. Participants were recruited using a combination of convenience and snowball sampling through the authors' extended personal and professional networks across the four regions. To ensure demographic diversity, a stratified sampling approach was applied to randomly select a balanced distribution of 100 participants across two distinct age groups: 18–25 years and older than 25 years. The dataset has a nearly equal class distribution with 1515 Hate and 1503 Non-Hate utterances, respectively. The distribution of speakers across regions, along with the variety of speakers and a balanced class distribution, makes AudioHate-DB a unique database for hate speech detection in the context of Bengali-accented English conversation.

All audio files are stored in uncompressed 16-bit PCM WAV format at a nominal sampling rate of 44,100 Hz (44.1 kHz), which is the same sampling rate as a normal speech-corpus file. The recordings were made in typical indoor and outdoor settings, using commercially available Android and iOS smartphones, thereby capturing a realistic amount of acoustic variability (device channel variations, reverberation, and ambient noise), typical of the real-world conditions a hate speech detection system would need to function in. Data acquisition parameters are summarized in (Table 2).

Table 2.

Audio data acquisition parameters.

Parameter	Specification
Recording Device	Mobile telephone (Android/iOS)
Audio Format	WAV (PCM)
Nominal Sampling Rate	44,100 Hz
Bit Depth	16-bit
Total Utterances	> 3000
Class Distribution	Balanced (Hate: ∼50%, Normal: ∼50%)
Speaker Diversity	Multi-regional, multi-gender
Acoustic Environment	Natural indoor/outdoor settings

The dataset contains the raw recordings and a version of each recording in which stationary background noise has been reduced using spectral-gating-based noise reduction. Processed files retain the same sampling rate, bit depth, and file naming as the original files, and each processed file is a direct duplicate of its corresponding noise-reduced file.

4. Experimental Design, Materials and Methods

The pipeline that was used to build the AudioHate-DB spoken by Bangladeshi speakers is shown in (Fig. 1). A corpus of 1006 pairs of sentences has been built, with almost 100 volunteer speakers recording the sentences on their smartphones in natural settings across 4 regions, from a curated hate lexicon that has 23 words and 6 categories. This resulted in 3018 WAV files that were denoised using spectral gating to produce a balanced processed dataset. Each step is described in detail in this section.

Fig. 1.

Dataset construction and pre-processing pipeline.

AudioHate-DB is built using a controlled elicitation design that aims to aid in the binary classification of hate speech in spoken Bangladeshi-accented English via audio signals.

4.1. Sentence corpus construction

The controlled sentence corpus contains 1006 sentences, including 505 sentences in the Hate class and 501 sentences in the Non-Hate class with a nearly balanced binary distribution. Sentences containing hate words were built using a hate lexicon of 23 pre-defined hate words. But, to give each sentence the Hate label, it was necessary that a lexicon term was present in the sentence as well as the overall context and meaning of it. A few members of the research group undertook human evaluation and the resulting Hate/Non-Hate labels were reviewed and validated. Lexicon was segmented into 6 categories of slurs: racial slurs, gendered and sexual slurs, ability-derogatory terms, LGBTQ+ slurs, general profanity, compound or contextual derogatory terms. (Table 3) presents representative examples of the Hate and Non-Hate utterances in the AudioHate-DB corpus, which were obtained after semantic and contextual analysis and human validation.

Table 3.

Representative utterances from the AudioHate-DB Corpus.

Category	Utterance (Text)	Label
Hate Speech	I'm a bitch, but guess what? You're still a slut.	Hate
Hate Speech	Eventually every bitch turns into that whore.	Hate
Hate Speech	Niggas is hoes, man, I swear.	Hate
Non-Hate Speech	I don't trust you.	Non-Hate
Non-Hate Speech	Don't really have anyone to talk to.	Non-Hate
Non-Hate Speech	Nobody cares about me.	Non-Hate

These examples show that offensive words are not the only basis for assigning a Hate label, but instead the meaning and context of the statements

4.2. Participants and recording protocol

Around 100 volunteer participants of different age groups, genders, and regional Bangladeshi accent backgrounds (Dhaka, Dinajpur, Cumilla, and Khulna) were recruited. The recruitment explicitly aimed to achieve regional and demographic diversity to capture naturally occurring phonetic and prosodic variation in the spoken English of the target language community (Bangladeshi-accented), thereby enhancing the representativeness of the data for the target language community.

Sentence sheets were prepared for each participant, who were asked to read each sentence aloud in a natural, unrehearsed voice, without acting or any intentional change to their speaking style. This approach was used to keep the recordings as close as possible to real conversational or social-media speech, even though the sentences themselves were predefined rather than freely spoken. Each sentence was read by multiple speakers, which introduces natural cross-speaker variation and supports the development of models that generalize across different speakers. All participants gave informed consent before recording and were informed in advance about the offensive nature of the hate-class material and the purpose for which the corpus was being collected.

4.3. Acoustic conditions and recording equipment

Commercial mobile telephones (Android and iOS smartphones) were all used for recording purposes. This was done on purpose because the mobile phone is the main communication and content creation tool in Bangladesh and similar contexts in South Asia, and recordings on a mobile phone will thus reflect the acoustic channel characteristics of the domain for which the dataset was collected. Audio was recorded at a 44,100 Hz (44.1 kHz) nominal sampling rate in uncompressed 16-bit PCM WAV format. Instead of acoustically treated studios, the recordings were made in natural indoor and outdoor spaces to include a realistic number of environmental variations (ambient noise, reverberation, and channel variations). The resulting corpus consists of 3018 utterances but is roughly balanced in classes (∼50% hate, ∼50% non-hate).

4.4. Audio preprocessing: noise reduction

All raw WAV files were first subjected to a noise reduction stage, as the recordings were made in uncontrolled indoor and outdoor environments. Noise reduction was applied to all audio recordings using the Noisereduce library. The default settings of the library were used without any manual tuning. The noise reduction was configured for non-stationary background noise, with a noise reduction proportion of 1.0. The spectral analysis used an FFT size of 1024, a window length of 1024 samples, and a hop length of 256 samples. The noise level was estimated using a 2-second time period. Frequency smoothing was set to 500 Hz, while time smoothing was set to 50 milliseconds. A threshold multiplier of 2.0 and a sigmoid slope of 10.0 were used for non-stationary noise estimation. All these values were the library's default settings and were not manually optimized in this study. This noise-reduced waveform was then saved to disk by the soundfile library (0.14.0), and the file kept the same class structure in /Processed Data/Hate/ and /Processed Data/Non_Hate/, respectively, as shown in (Fig. 2).

Fig. 2.

Directory structure of the AudioHate-DB dataset.

This pre-processing stage is used to reduce the stationary background noise created by the natural indoor/outdoor acoustic environments, which results in better quality and discriminability of the acoustic features extracted in the following stages and improves the signal-to-noise ratio of the recordings. The processed corpus is identical to the original with respect to sampling rate, bit depth, and file names to ensure that each recording in the corpus was matched with its noise-reduced version.

4.5. Acoustic feature extraction

All processed WAV files are processed through feature extraction. Chroma Shifting: Aggregated the 12-D pitch class profile to a single scalar. Root Mean Square energy: Mean of the energy of each frame. Spectral Centroid: geometric center of the sound spectrum. Weighted Standard Deviation of the Spectrum: the weighted standard deviation of the spectrum. Spectral Rolloff: Frequency below which 85% of the energy in the spectrum lies. Zero Crossing Rate: Mean value of the zero-crossing rate per frame. Jitter, MFCCs: The first 13 MFCC coefficients are averaged over frames to give 13 scalar features [10].

This results in a 20-dimensional feature vector per audio sample (1 chroma + 1 RMS + 1 centroid + 1 bandwidth + 1 rolloff + 1 ZCR + 1 jitter + 13 MFCCs) with the class label appended. All the feature vectors are written in an incremental fashion into a CSV file that has a predefined header. Before model training, the dataset is checked for null values that were not seen in the dataset. (Table 4) provides a complete description of each feature, its acoustic domain, and its relevance to speech classification.

Table 4.

Summary of dataset features.

Feature	Description	Equations
Chroma STFT Mean	STFT-based mean energy across 12 pitch classes, capturing harmonic and tonal characteristics.	$\begin{array}{c}{X}_{PCP}\left(\mathrm{K}{}^{\prime },\mathrm{n}\right)=\sum _{K:P\left(K\right)=K^{\prime }}{X}_{STFT}\left(K,n\right)\left(1\right)\end{array}$
RMS Mean	Average of the square root of the amplitude of a data frame; measures loudness and intensity of a signal.	$\begin{array}{c}RM{S}_{mean}=\frac{1}{T}\sum _{t=1}^T\sqrt{\frac{1}{n}}\sum _{i=1}^N{x}_i^2\left(t\right)\left(2\right)\end{array}$
Spectral Centroid Mean	Power-spectrum weighted mean frequency, indicating perceived brightness.	$\begin{array}{c}S{C}_{mean}=\frac{1}{T}\sum _{t=1}^T\frac{{\sum }_{k=1}^K{f}_k{S}_{k,t}}{{\sum }_{k=1}^K{S}_{k,t}}\left(3\right)\end{array}$
Spectral Bandwidth Mean	Weighted spectral standard deviation around the centroid, measuring spectral spread.	$\begin{array}{c}S{B}_{mean}=\frac{1}{T}\sum _{t=1}^T\sqrt{\frac{{\sum }_{k=1}^K{S}_{k,t}{\left(f_k-C_t\right)}^2}{{\sum }_{k=1}^K{S}_{k,t}}}\left(4\right)\end{array}$
Spectral Rolloff Mean	85% spectral energy frequency, distinguishing harmonic content from noise.	$\begin{array}{c}S{R}_{mean}=\frac{1}{T}\sum _{t=1}^T\frac{f_s}{2\left(K-1\right)}r,t\left(5\right)\end{array}$
Zero Crossing Rate Mean	Mean zero-crossing rate per frame, indicating noisiness and fundamental frequency.	$\begin{array}{c}ZC{R}_{mean}=\frac{1}{T}\sum _{t=1}^T\frac{f_s}{N}\sum _{i=1}^{N-1}\left|p\left(x\left[i\right]\right)-p\left(x\left[i-1\right]\right)\right|\left(6\right)\end{array}$
Jitter Mean	Mean cycle-to-cycle F0 variation, indicating vocal tension and stress.	$\begin{array}{c}Jitter=\frac{\frac{1}{N_p-1}{\sum }_{i-1}^{N_p-1}\left|P_i-P_{i+1}\right|}{\frac{1}{N_p}{\sum }_{i=1}^{N_p}{P}_i}\left(7\right)\end{array}$
MFCC 1–13 Mean	Mean of the first 13 MFCCs, capturing spectral envelope and vocal tract characteristics.	$MFCC\left(i,l\right)=$ $\begin{array}{c}{\alpha }_i\sum _{m=1}^{M}10lo{g}_{10}\left[\sum _k{H}_m\left(k\right)S\left(k,l\right)\right]\text{cos}\left[\frac{\pi i\left(m-\frac{1}{2}\right)}{M}\right]\left(8\right)\end{array}$ $wherei=0,1,\dots ..,12$

5. Model Development and Experimental Results

5.1. Data partitioning

The 3018 utterances were divided into non-overlapping training (65%), validation (15%), and test (20%) sets using stratified sampling. The same random seed was used for both splitting stages to ensure reproducibility.

Five-fold stratified cross-validation was used on the training set. The 604-utterance test set was reserved for final evaluation, while the validation set was used for LSTM early stopping. All 20 descriptors were standardized to zero mean and unit variance using only the respective training data to prevent data leakage.

5.2. Traditional machine learning classifiers

Three of the standard supervised classifiers were implemented using pipelines with feature standardization to avoid data leakage in scikit-learn. None of the systems were tuned with hyperparameters and were treated as baseline systems.

Three classical classifiers were implemented using standardized features within a pipeline. Logistic Regression used L2 regularization (C = 1.0) with a maximum of 5000 iterations. Random Forest was configured with 100 decision trees and parallel processing enabled. Gradient Boosting used 100 estimators, a learning rate of 0.1, and a maximum tree depth of 3. All models used a fixed random seed (42) to ensure reproducibility

5.3. Long short-term memory network

A neural baseline was created using a deep learning-based LSTM. The representation for each utterance was a 20-dimensional feature vector and the utterance was considered as a sequence of length one. Therefore, the LSTM is able to model non-linear feature interactions and does not model the temporal dynamics, leaving this area for future research.

The LSTM network was composed of a 128-unit recurrent layer, a 0.2 dropout rate layer and a 2-unit output softmax activation layer. The model was trained with Adam optimizer, learning rate of 0.001, batch of 64, maximum epochs of 50 with categorical cross-entropy loss. Early stopping was used with patience of 5, based on the validation loss and the best weights were recovered. The five-fold stratified cross-validation was used with the fitting of the scaling parameters conducted within each fold separately to avoid leakage.

5.4. Traditional classifier performance

The performance of the model was measured with accuracy and the area under the receiver operating characteristic curve (AUC-ROC). Five-fold cross-validation was performed only on the training set (n = 1961), and the final performance was measured on an independent held-out test set (n = 604). The results obtained are summarized in (Table 5).

Table 5.

Overall performance metrics of traditional classifiers.

Classifier	Test Accuracy	CV Mean Acc. (±SD)	AUC-ROC
Logistic Regression	70.53%	72.87% (±2.11%)	0.7832
Random Forest	82.78%	84.04% (±1.33%)	0.9159
Gradient Boosting	79.64%	82.97% (±1.31%)	0.8928

5.5. LSTM network performance

The LSTM used Adam (learning rate = 0.001), 50 epochs, batch size 64, and categorical cross-entropy. Early stopping monitored validation loss (patience = 5), while the test set remained unseen. Training reached the full 50 epochs, achieving 88.08% validation and 83.94% test accuracy.

5.6. Comprehensive model performance evaluation

As shown in (Fig. 3), the ROC curves on the held-out test set (n = 604) showed that all models outperformed random classification. LSTM achieved the highest AUC (0.9264), followed by Random Forest (0.9159), Gradient Boosting (0.8928), and Logistic Regression (0.7832). LSTM and Random Forest performed comparably, while Logistic Regression showed substantially weaker discrimination.

Fig. 3.

ROC Curve analysis of the four classification models.

5.7. Comparative model performance

Comprehensive comparison of four machine learning and deep learning classifiers across different evaluation metrics are reflected in (Table 6).

Table 6.

Class-wise performance comparison of the evaluated classification models.

Model	Class	Precision	Recall	F1	Acc.	Errors
Logistic Regression	Hate	0.7010	0.7195	0.7101	70.53%	178
	Non-Hate	0.7099	0.6910	0.7003
Random Forest	Hate	0.8443	0.8053	0.8243	82.78%	104
	Non-Hate	0.8127	0.8505	0.8312
Gradient Boosting	Hate	0.8261	0.7525	0.7876	79.64%	123
	Non-Hate	0.7713	0.8405	0.8045
LSTM	Hate	0.8503	0.8251	0.8375	83.94%	97
	Non-Hate	0.8595	0.8538	0.8567

The comparison reveals that the LSTM model recorded the highest accuracy of 83.94% and the least number of errors of 97 among all the models, showing the best classification performance. Random Forest had 82.78% accuracy and 104 errors, the second best, followed by Gradient Boosting with 79.64% accuracy and 123 errors, and Logistic Regression with 70.53% accuracy and 178 errors.

Limitations

AudioHate-DB is limited to a relatively small number of volunteer speakers and has been recorded in four districts of Bangladesh, which may not be representative of the country's regional accent variations. In particular, accents of Chittagong and Chapainawabganj are missing. In the future, with more regional accents from Bangladesh and with speakers from other countries in South Asia, could be expanded to increase the diversity and applicability of the dataset. The recordings were anonymized at the time of capture, which means that we do not have a speaker identifier on the recordings and therefore cannot create a speaker-disjoint partition. Words can then be duplicated in the training and testing sets. This is important because the acoustic features used here are not only correlated with content being hateful vs. non-hateful but also correlated with speaker identity; thus, some of the test error reduction achieved by a classifier can come from recognizing a familiar speaker, rather than learning the acoustic features of hate speech. The reported figures are therefore speaker-dependent and speak to an upper limit on accuracy that is achievable for unseen speakers. They are reported on the released partition for ease of future reproduction and comparison.

CRediT Author Statement

Arif Faisal: Conceptualization, Data curation, Methodology, Software, Formal analysis, Writing – Original draft, and Visualization. Syeda Sumaia Sultana: Data curation, Methodology, Validation, Investigation; Writing-Review and Editing, and Visualization. Suprove Chandra Sarkar: Conceptualization, Supervision, Writing-Review, and Editing. Shabnom Mustary: Project management, Supervision, and Writing-Review.

Ethics Statement

This research protocol was reviewed and formally approved by the Ethical Review Committee (ERC) of the Division of Research (DoR) at Daffodil International University, Dhaka, Bangladesh (Ref: ERB/MKH/202625108, approved August 25, 2026). Prior to data collection, all participants underwent age verification to confirm they were at least 18 years of age. Participants were explicitly briefed on the research objectives and given prior notification regarding the presence of offensive, profane, and derogatory terms within the reading prompts. Participation was strictly voluntary, and written/verbal informed consent was obtained before recording. An explicit opt-out protocol was maintained: participants retained the right to skip specific prompt sentences or withdraw from the study entirely at any point without penalty or negative consequences. To protect confidentiality, no personally identifiable information was collected or linked to audio files. Recordings were assigned unique alphanumeric codes, anonymized at the point of capture, and stripped of personal metadata. All datasets are used strictly for academic and educational speech processing research under the approved protocol.