niloycste68/Bangali_local_dialect_ASR_HF_Dataset

BanglaMix — Code-Switching ASR in Bangladeshi Regional Dialects

BanglaMix is a speech dataset for Automatic Speech Recognition (ASR) on dialectal Bangladeshi Bengali mixed with English (code-switching). It covers 15 regional dialects and the natural Bengali–English code-switching common in informal Bangladeshi speech — a setting not covered by existing Bengali corpora, which address either dialects or code-switching, never both.

Clips	41,499 transcribed audio clips
Duration	~131.2 hours
Splits	train 34,605 · validation 2,806 · test 4,088
Dialects	15 Bangladeshi regional dialects
Languages	Bengali (bn) + English (en) code-switching
Audio	16 kHz mono, 5–15 s per clip
Code-switched clips	2,361 (~5.7%, naturally occurring)
License	CC BY-NC 4.0 (non-commercial research use)

These 41,499 clips are the silver-standard–validated subset of 51,857 segmented clips. This release contains real recordings only — audio augmentation used for training the paper's models is a training-time choice and is not included; apply it yourself if needed.

⚠️ Labels are silver-standard — transcripts were generated by Whisper large-v3 and validated by an LLM judge, not by human transcribers. They contain label noise (dialectal Bengali also has no standardized spelling). Use CER as the primary metric and treat the dataset as a challenging open benchmark, not gold-standard ground truth.

---

How to use

from datasets import load_dataset, Audio

ds = load_dataset("niloycste68/Bangali_local_dialect_ASR_HF_Dataset")

# decode audio to 16 kHz waveforms on access
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

example = ds["train"][0]
print(example["sentence"])           # transcript (Bengali, may contain English)
print(example["dialect"])            # e.g. "Sylhet"
print(example["is_code_switched"])   # True / False
waveform = example["audio"]["array"] # numpy float array @ 16 kHz

Filter to only code-switched clips:

cs = ds["train"].filter(lambda x: x["is_code_switched"])

---

Data fields

Field	Type	Description
audio	Audio (16 kHz)	The speech clip
sentence	string	Transcript (Bengali script; may contain Latin-script English words)
clip_id	string	Unique clip identifier
dialect	string	One of the 15 regional dialects
domain	string	Content domain (e.g. vlog, tech review, street interview)
is_code_switched	bool	Whether the clip contains Bengali–English code-switching
bn_ratio	float	Fraction of Bengali-script words
en_ratio	float	Fraction of English-script words
duration_sec	float	Clip duration in seconds

---

Dialect coverage

Dialect	Clips	Hours
Sylhet	7,931	26.0
Old Dhaka	4,741	14.6
Sandwip	4,583	13.0
Barishal	3,259	11.0
Chittagong	2,993	9.1
Khulna	2,774	8.9
Narail	2,371	7.8
Kishoreganj	2,287	7.1
Rangpur	1,985	6.4
Comilla	1,847	6.1
Narsingdi	1,661	5.1
Habiganj	1,446	4.5
Mymensingh	1,369	4.1
Noakhali	1,140	3.6
Tangail	1,112	3.8
Total	41,499	131.2

Distribution follows online content availability and is intentionally not a demographic sample; the imbalance is reported transparently so users can re-weight or resample as needed.

---

How it was built

Audio was collected from public social-media content using dialect-name search queries (which surface creators who deliberately use the dialect) across code-switch-rich genres (vlog, technology/phone review, campus/career, street interview, gaming). Long recordings were segmented into 5–15 s clips with WebRTC voice-activity detection, transcribed with Whisper large-v3, and labelled for code-switching from Unicode script ranges (Bengali vs. Latin). High-confidence transcripts were auto-accepted; borderline ones were validated by an LLM judge (Qwen2.5). No human transcribers were used.

---

Intended use & limitations

• Intended use: training and evaluating ASR systems for dialectal and code-switched Bengali; research on low-resource and code-switching speech.
• Limitations: silver-standard labels (label noise); no standardized dialectal orthography (word-level WER over-penalizes valid spelling variants — prefer CER); imbalanced dialect coverage; social-media domain (may not transfer to broadcast/clinical speech). Because labels come from Whisper, evaluations may slightly favour Whisper-family models.
• Ethics: sourced from publicly available content; no speaker identities, demographics, or contact information are collected or released. Non-commercial research use only.

---

Citation

If you use BanglaMix, please cite the accompanying paper. Code and full reproduction pipeline: https://github.com/niloycste/ASR-Bengali-Local-Dialect-

@misc{islam2026banglamix,
  title  = {BanglaMix: A Dataset and Benchmark for Code-Switching ASR in Bangladeshi Regional Dialects},
  author = {Islam, M. Mohaiminul},
  year   = {2026},
  note   = {Dataset: https://huggingface.co/datasets/niloycste68/Bangali_local_dialect_ASR_HF_Dataset}
}