• Developed by: Researcher at Dept. of CSE, SUST, Bangladesh
• Paper: BPAC: How Detectable is Regional Accent in Read Bangla? A Parallel Corpus and its Phonetic Structure (in preparation)
• Point of Contact: Prof. Dr. M. Shahidur Rahman, Dept. of CSE, SUST

BPAC: Bangla Parallel Accent Corpus (Multi-Rate Audio Release)

Status: companion research paper in preparation, not yet submitted/published. This dataset is released ahead of its companion manuscript, "-------------------------------------------------" Please cite the dataset directly (see Citation below) until the paper is published, this card will be updated with the full citation at that point.

Dataset summary

BPAC is a parallel, regional-accent speech corpus for Bangla (Bengali): 51 native speakers from 8 regions of Bangladesh; Sylhet, Chattogram, Dhaka, Faridpur, Khulna, Mymensingh, Rajshahi, and Rangpur; each read an identical, phonetically diverse 250-sentence script, so that region and speaker vary while linguistic content is held fixed by design. The result is 12,747 utterances (three recordings were excluded as unusably short or silent) across all speakers and regions, with every sentence available in every region for direct, content-controlled comparison.

This parallel design is what distinguishes BPAC from every other public Bangla speech corpus we are aware of: it is the only one, to our knowledge, in which the same text is read in full by a defined sample of speakers across multiple named regions, making it usable for accent detectability research, accent/voice conversion (a same-content reference exists for every utterance), and accent-controlled ASR evaluation, not only general-purpose ASR acoustic modeling.

• Speakers: 51 (25 female, 26 male), aged 19–25
• Regions: 8 (Sylhet, Faridpur, Khulna, Chattogram, Rangpur, Mymensingh, Rajshahi, Dhaka)
• Script: 250 sentences, identical across every speaker
• Utterances: 12,747
• Audio tiers: original 48 kHz stereo (lab) / 44.1 kHz mono (app), plus a 16 kHz mono working corpus (peak-normalized to −1 dBFS) — see Audio format
• Total audio: ~27.7 hours (per tier)
• Language: Bangla / Bengali (bn)
• Recording pipelines: laboratory (male speakers, studio microphone) and a purpose-built Android app, SUST Kotha (female speakers, raw unprocessed microphone signal; no AGC, noise suppression, or echo cancellation applied)

Dataset structure

BPAC/
├── data/                          16 kHz mono working corpus (peak -1 dBFS)
│   ├── metadata.csv
│   ├── chattogram/
│   │   ├── chattogram_female_001_001.wav
│   │   ├── chattogram_female_001_002.wav
│   │   └── ...
│   ├── dhaka/
│   ├── faridpur/
│   ├── khulna/
│   ├── mymensingh/
│   ├── rajshahi/
│   ├── rangpur/
│   └── sylhet/
└── data_original/                 original, un-normalized audio (native
    ├── metadata.csv                sample rate/channels per pipeline)
    ├── chattogram/
    ├── dhaka/
    ├── faridpur/
    ├── khulna/
    ├── mymensingh/
    ├── rajshahi/
    ├── rangpur/
    └── sylhet/

Both trees mirror the same per-region, per-utterance filenames; only the root folder and the underlying audio differ. Each folder carries its own metadata.csv, so an utterance can be retrieved from either tier by the same relative path.

data/metadata.csv and data_original/metadata.csv are structurally identical; one row per utterance, the same eight columns, the same values. Only the audio behind file_name differs:

Column	Description
file_name	Path to the audio file, relative to the folder holding this metadata file (e.g. chattogram/chattogram_female_001_001.wav). Decoded automatically by datasets into an audio column.
utterance_id	Anonymised unique utterance identifier: <region>_<gender>_<speaker_no>_<sentence_id>
speaker_id	Anonymised unique speaker identifier: <region>_<gender>_<speaker_no>
region	One of the 8 regions listed above
gender	male / female
pipeline	lab or app; the recording pipeline (see note on confound below)
sentence_id	Integer 1–250, identifying which script sentence this is; shared across all speakers and regions
sentence	The Bangla sentence text read in this recording

Because sentence_id is shared across every speaker, filtering on it returns all 51 renditions of the same sentence, the core parallel-corpus operation. See How to load.

Composition by region

Every speaker read all 250 sentences. There are no missing (speaker, sentence) cells beyond the three excluded empty files.

Region	Male	Female	Speakers	Utterances	Audio (min)	Mean utt. (s)
Sylhet	6	4	10	2,500	297.2	7.13
Faridpur	4	3	7	1,749	231.1	7.92
Dhaka	3	3	6	1,500	210.2	8.41
Chattogram	3	3	6	1,500	206.7	8.27
Khulna	3	3	6	1,500	200.3	8.01
Rajshahi	3	3	6	1,500	190.3	7.61
Mymensingh	2	3	5	1,248	167.0	8.02
Rangpur	2	3	5	1,250	161.6	7.76
Total	26	25	51	12,747	1,664.5	7.83

Important confound to be aware of: recording pipeline is perfectly confounded with gender in this corpus. Every lab recording is male, every app recording is female. No claim about pipeline/channel effects can be separated from gender effects using this corpus alone; any apparent "pipeline effect" is equally validly read as a "gender effect," and vice versa.

How to load

from datasets import load_dataset
 
# 16 kHz working corpus (default)
ds = load_dataset("ArfatulAsif/BPAC")
 
# original-rate recordings
ds_orig = load_dataset("ArfatulAsif/BPAC", "original")
 
print(ds["train"][0])

Both configs decode audio automatically, so ds["train"][0]["audio"] returns a dict containing the decoded waveform array and its sampling rate, alongside every metadata field.

To retrieve every speaker's recording of a specific sentence the core parallel-corpus operation:

sentence_42 = ds["train"].filter(lambda x: x["sentence_id"] == 42)

To work within a single region or recording pipeline:

sylhet = ds["train"].filter(lambda x: x["region"] == "sylhet")
lab_only = ds["train"].filter(lambda x: x["pipeline"] == "lab")

Audio format

This release ships two audio tiers for every utterance:

• data_original/: the original, un-normalized recordings at their native sample rate and channel count per recording pipeline: 48 kHz, stereo for the laboratory pipeline (male speakers), 44.1 kHz, mono for the mobile-app pipeline (female speakers). No resampling, channel-mixing, or level normalization has been applied, this is the signal as captured, minus the anonymization rename.
• data/: a 16 kHz, mono, 16-bit PCM WAV working corpus, peak-normalized to −1 dBFS, uniform across both pipelines. This is the tier metadata.csv's file_name column points to, and the one the datasets audiofolder loader will decode by default.

Use data_original/ (via file_name_original) for analyses sensitive to sample rate or channel count: for example, sibilant/fricative spectral measurements, where a 16 kHz tier's 8 kHz Nyquist limit clips exactly the upper-frequency energy such measurements depend on. Use data/ for general-purpose modeling where a uniform, normalized format across both recording pipelines matters more than native fidelity.

Collection

In brief: participants were recruited as native speakers of their stated region, aged 19–25, mostly university students. Male participants were recorded in a sound-treated laboratory (SUST Audio Laboratory) via a dedicated audio interface; female participants recorded independently via a purpose-built Android application after a remote training session on minimizing background noise. All participants gave informed consent (written form, and equivalent Google Form for lab participants; Google Form for app participants) and received compensation; consent explicitly authorized public release of voice recordings together with region, gender, and age, while withholding names and contact information from any public release.

Known corpus corrections

One systematic labeling error, affecting 29 of 12,750 raw utterances for a single speaker, was found during an internal verification audit and corrected prior to this release (a recording/filing rotation in which each affected file's audio content belonged to the next sentence in a contiguous range). Full detail is documented in the companion manuscript's methodology section. The dataset as released here reflects the corrected, fully verified state, every one of the 12,747 included utterances has been confirmed to contain its labeled sentence, in both audio tiers.

Intended uses

• Bangla accent/dialect detection and classification research
• Accent-controlled or accent-robust ASR training and evaluation
• Accent conversion / voice conversion (parallel, same-content reference available for every utterance)
• Acoustic-phonetic dialectometry (vowel formants, rhythm, sibilant realization, intonation, forced-alignment-based duration comparison) — use data_original/ for measurements sensitive to sample rate or channel count
• General-purpose Bangla ASR acoustic modeling

License

This dataset is released under CC BY-NC 4.0 (Attribution-NonCommercial 4.0 International). You are free to share and adapt this material for any non-commercial purpose, with appropriate attribution. See https://creativecommons.org/licenses/by-nc/4.0/ for the full license text. For commercial licensing inquiries, contact the authors (below).

Citation

A peer-reviewed paper describing this corpus and its phonetic analysis is currently in preparation and has not yet been submitted or published. Until it is available, please cite the dataset directly:

@dataset{bpac2026,
  title        = {BPAC: Bangla Parallel Accent Corpus},
  author       = {Asif, Arfatul Islam and Khan, Unayes Ahmed and Joy, Sanjoy Das and Rahman, Mohammad Shahidur},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ArfatulAsif/BPAC}},
  doi          = {https://doi.org/10.57967/hf/10299},
  note         = {Dataset accompanying a manuscript in preparation,
                  "-------------------------------------------------." Cite this
                  dataset directly until the companion paper is published.}
}

Contact

Department of Computer Science and Engineering, Shahjalal University of Science and Technology (SUST), Sylhet, Bangladesh. Research team: Arfatul Islam Asif, Unayes Ahmed Khan, Sanjoy Das Joy. Supervisor: Professor Mohammad Shahidur Rahman, PhD.

✎ Contact email: rahmanms@sust.edu (Professor Mohammad Shahidur Rahman, PhD)