Bangla-10K: A Challenging, Metadata-Rich Corpus of Read and Conversational Bengali Speech from India and Bangladesh

Bangla-10K is a 10,816-hour Bengali speech corpus with 624,951 recordings from India and Bangladesh: a 10,070.8-hour core corpus (567,323 recordings) and a separately collected 745.1-hour evaluation set (57,628 recordings). It combines scripted single-speaker read speech with natural multi-speaker conversations for Bengali automatic speech recognition (ASR).

The corpus and its ASR evaluation are described in the manuscript Bangla-10K: A Challenging, Metadata-Rich Corpus of Read and Conversational Bengali Speech from India and Bangladesh.

Unless explicitly labeled as paper-reported, file formats, storage layout, topic-labeling details, and release metadata below describe the current Hub artifact rather than claims made in the paper.

Core corpus (single_speaker + multi_speaker)	10,071 hours (567,323 recordings)
Open release (preview)	849.4 hours (62,573 recordings): the full 745.1-hour evaluation set plus a 104-hour sample of the core corpus, CC BY 4.0
Read speech (single_speaker)	6,537 h · 500,018 recordings
Conversation (multi_speaker)	3,534 h · 67,305 recordings · 3 tracks each (isolated speaker tracks plus combined mix)
Speaker metadata	age, gender, country, region on every recording
Transcripts	reading prompts for read speech; human-transcribed, speaker-attributed segments for conversations
Audio	WAV (PCM or IEEE float) · 44.1 or 48 kHz · mono or stereo · 16, 24, or 32 bit
Splits	train / validation / test with no speaker overlap
Integrity	SHA256 published for every audio object

Highlights

• Scale with provenance. Every recording was commissioned for this corpus rather than scraped from the web; demographic metadata accompanies every recording and every audio byte is checksummed.
• Two registers. Scripted single-speaker readings and natural conversation between two speakers, with isolated tracks per speaker plus the combined mix for diarization, separation, and turn-taking research.
• Real world Bengali. bn-IN and bn-BD varieties with the natural English code switching of contemporary speech ("credit card", "EMI", and so on).
• Quality gated. Automated checks validate metadata, lexical diversity, audio quality, linguistic accuracy, and audio-transcript alignment. Native Bengali linguists review stratified samples, set quality cutoffs, and return poor transcripts for revision. Human review is sample based, not a manual review of every recording.
• Benchmark ready splits. Train, validation, and test are disjoint at the speaker level, across both subsets, assigned deterministically.

Quick start

from datasets import load_dataset

# This repository is gated: accept the access terms on the Hub, then authenticate before loading.
# Load the 849-hour open release (audio embedded).
preview = load_dataset("psdn-ai/bangla-10k", "preview", split="sample", token=True)
preview[0]["audio"]          # decoded audio
preview[0]["transcript"]     # full transcript

# Full metadata and transcripts (audio lives in external storage, see below)
single = load_dataset("psdn-ai/bangla-10k", "single_speaker", token=True)
multi = load_dataset("psdn-ai/bangla-10k", "multi_speaker", token=True)

# Stream and filter without downloading everything
ds = load_dataset("psdn-ai/bangla-10k", "single_speaker", split="train", streaming=True, token=True)
bd_banking = ds.filter(lambda r: r["locale"] == "bn-BD" and r["topic"] == "Banking")

Full resolution audio (about 6.6 TB) is hosted in a private Cloudflare R2 bucket with free egress; each row's audio_files[].object_key and sha256 locate and verify every file. See docs/download_audio.md.

Subsets

Subset	Rows	What it is	Audio
preview	62,573 recordings (849.4 h)	open release: the 745.1-hour evaluation set plus a 104-hour core sample; CC BY 4.0	embedded
single_speaker	500,018	one row per read-speech recording	external (R2)
multi_speaker	67,305	one row per conversation, 3 audio tracks	external (R2)

The core-corpus sample in the preview is selected from the train split for audio/transcript quality, speaker and topic coverage, and lexical diversity; the evaluation set is included in full. Conversation previews use the combined mix. Audio is embedded as lossless FLAC for in-browser playback; the full corpus is delivered as the original WAV files.

Dataset structure

Example instance (single_speaker, truncated)

{
  "audio_id": "001ddde7-49b0-4437-b8e5-0931ff6855a9",
  "split": "test",
  "language": "bn",
  "locale": "bn-IN",
  "topic": "Current Events",
  "topic_source": "vendor",
  "duration": 67.66899871826172,
  "sample_rate": 44100,
  "channels": 1,
  "format": "wav",
  "transcript_key": "v1/single_speaker/transcripts/001ddde7-49b0-4437-b8e5-0931ff6855a9.json",
  "segments": [
    {
      "speaker": "speaker_a",
      "start_seconds": 0.0,
      "end_seconds": 68.0,
      "text": "সেদিন office থেকে ফেরার সময় যা একটা কাণ্ড হলো না! বৃষ্টি তো…",
      "raw_text": null
    }
  ],
  "speakers": [
    {
      "speaker": "speaker_a",
      "speaker_id": "ccf9b404-2dd7-445f-bbfa-572d947c53ee",
      "age": 29,
      "gender": "female",
      "country": "India",
      "region": "Kolkata"
    }
  ],
  "audio_files": [
    {
      "track": "speaker_a",
      "object_key": "v1/single_speaker/audio/001ddde7-49b0-4437-b8e5-0931ff6855a9.wav",
      "sha256": "71e921ccc0ed499e6f25df1795fa8ff21cf45e4f79762ac58a0c7cf54044f3b9"
    }
  ]
}

Data fields (single_speaker / multi_speaker)

Field	Type	Notes
audio_id	string	Public UUID, unique per subset
split	string	train / validation / test
language	string	Always bn
locale	string	bn-BD or bn-IN
topic	string	One of 8 topics (below)
topic_source	string	vendor (collection label) or model (LLM assigned)
duration	float32	Seconds, from the WAV header
sample_rate	int32	44100 or 48000
channels	int8	1 (mono) or 2 (stereo)
format	string	wav
transcript_key	string	Object key of the standalone transcript JSON
segments	list⟨struct⟩	speaker, start_seconds, end_seconds, text, raw_text
speakers	list⟨struct⟩	speaker, speaker_id, age, gender, country, region
audio_files	list⟨struct⟩	track, object_key, sha256

audio_files[].track is speaker_a for read speech; conversations carry speaker_a and speaker_b (isolated) and combined. speakers[].speaker_id is stable for the same speaker profile across recordings. segments[].raw_text is null where identical to text.

Preview fields: audio, audio_id, mode, locale, topic, duration, transcript, speaker_genders, speaker_ages, speaker_countries. No field is ever empty.

Splits

No speaker appears in more than one split, including across subsets (splits are assigned over connected components of speakers who appear together).

	train	validation	test
single_speaker recordings	494,971	2,468	2,579
single_speaker hours	6,471.16	32.27	33.82
multi_speaker recordings	66,634	345	326
multi_speaker hours	3,498.88	18.12	16.57

Topics

Topic	Hours
Business	2,178
Social Topics	1,503
Sports	1,277
Politics	1,113
Customs of Target Regions	1,064
Personal Finance	1,044
Banking	989
Current Events	904

Topic labels come from collection prompts where available; 19% of rows (mostly conversational batches) were labeled from their transcripts with Llama 3.3 70B (temperature 0, fixed taxonomy of 8 topics) and carry topic_source = "model". Treat those as high quality weak labels.

Speaker demographics

Speaker attributes are captured at collection time. Counts below are distinct pseudonymous speaker_id profiles reported by the release metadata, not the paper's per-recording speaker entries or a verified count of unique people. Hours count recordings featuring at least one speaker of the group.

Gender	Speaker profiles	Hours
female	387,002	6,967
male	169,382	3,959

Country	Speaker profiles	Hours
India	535,936	9,279
Bangladesh	20,448	792

Age band	Speaker profiles
17 to 25	268,541
26 to 35	227,062
36 to 45	47,100
46 to 60	13,006
61 and above	675

Locales (recordings): bn-IN 551,382 · bn-BD 15,941

The paper reports demographic shares over speaker entries, not unique speakers. Read speech is 76% female, 84% age 18-34, and 0.1% from Bangladesh; conversation is 45% female, 98% age 18-34, and 23% from Bangladesh. The paper identifies the concentration of young, female, and Indian speakers in the read subset as a limitation of the four-month recruitment window.

Dataset creation

Collection. Recorded remotely on consumer devices across India and Bangladesh through a commissioned collection program in sixteen delivery batches during a four-month collection window in 2026. Native Bengali speakers read LLM-generated prompts designed for lexical diversity. Separate multi-speaker sessions contain free-flow conversations on assigned topics, captured with isolated tracks per speaker plus a combined mix. Participants consented, were fairly compensated, and collection complied with applicable data rights and privacy laws.

Transcription. Read-speech transcripts are the corresponding prompts. Conversations were transcribed from audio by human transcribers, turn by turn, with speaker labels and start and end times. The transcripts preserve natural code-switching, overlap, and regional language variation.

Quality assurance. Every audio-transcript pair and metadata file passes automated metadata, lexical-diversity, and audio-quality checks. LLM judges score every sample for spelling, grammar, semantics, error severity, and audio-transcript alignment, including missing, extra, and mismatched words. Native Bengali linguists then inspect stratified samples selected from those scores using the same rules. Their feedback sets quality cutoffs, sends poor transcripts for revision, and recalibrates prompt generation and judging.

Paper evaluation

The paper fine-tunes a Bengali-pretrained Whisper-medium model on Bangla-10K and reports corpus-level CER and WER. On a 100-hour held-out Bangla-10K corpus set (16,744 clips) the full fine-tuned model reaches 10.87% CER and 16.45% WER, outperforming all evaluated open-source baselines. On a stratified 4,000-clip (51.5 h) subset of the independent 745.1-hour evaluation set it reaches 5.02% CER and 10.87% WER. See the manuscript for the full protocol, baselines, and commercial API comparisons. The fine-tuned checkpoint is released at psdn-ai/bangla-asr.

License and access

The paper reports two access tiers:

• The preview config (62,573 recordings, 849.4 hours: the full 745.1-hour evaluation set and a 104-hour sample of the core corpus) is released under CC BY 4.0.
• The full core corpus is available under a data-use agreement.

The repository is gated: accept the access terms on the Hub to download. The single_speaker and multi_speaker configs hold the metadata and transcripts of the full core corpus; their full-resolution audio requires R2 credentials issued under the data-use agreement. Request access or licensing.

Paper and dataset citation

The paper is under review. Until the publication record is available, cite:

@misc{nagpal2026bangla10k,
  title  = {{Bangla-10K}: A Challenging, Metadata-Rich Corpus of Read and Conversational Bengali Speech from India and Bangladesh},
  author = {Nagpal, Rashmi and Liu, Lefei and Rupaakula, Krishna and Shah, Sarick and Park, Jongwon and Chinchali, Sandeep P.},
  year   = {2026},
  note   = {Manuscript under review}
}

For the dataset artifact itself:

@misc{psdnai2026bangla10kdataset,
  title     = {{Bangla-10K} Dataset},
  author    = {{Poseidon AI}},
  year      = {2026},
  doi       = {10.57967/hf/10567},
  url       = {https://huggingface.co/datasets/psdn-ai/bangla-10k}
}

See also: docs/download_audio.md · docs/data_statement.md · docs/changelog.md

Batch 17 preview addition

The preview configuration now contains 62,573 recordings across 261 consistently numbered Parquet shards, representing approximately 849.4 hours of Bengali audio. Batch 17 contributes 57,628 single-speaker recordings and 745:05:35 of audio.

• India records use locale bn-IN; Bangladesh records use bn-BD.
• The 750 records outside India and Bangladesh use generic locale bn.
• Allowed gender values include female, male, and prefer-not-to-say.
• The topic label Customs of Target Regions is preserved unchanged.