bolAIndia

Human-side speech from production call recordings, cut into utterance-level chunks by a two-engine VAD (Silero + TEN) and transcribed by third-party ASR providers. Each row keeps the transcript, the provider's confidence, and full provenance back to the source recording.

Alongside it, open Indian-language speech corpora converted to the same schema (16 kHz mono FLAC, one utterance per row), each in a config of its own and tagged with where it came from.

Size

	hours	chunks	configs
Call-centre speech, vendor transcripts	20.56K	24.30M	3
Open Indian-language datasets	116.20K	43.43M	27
All	136.76K	67.73M	30

K = thousand, M = million. Exact figures are in STATS.md.

Repeats are removed: an open-dataset clip is published once, and call-centre rows that repeated another row's audio and transcript were removed on 5 October 2026. A call chunk that two vendors transcribed differently keeps one row per vendor.

Sources

One config per transcription system, so their output stays separable.

config (source_id)	provider	model	hours	chunks	shards
vendor-a	vendor-a	undisclosed	418	478K	42
combined	all vendors	mixed	20.1K	23.8M	4.6K
vendor-b	vendor-b	undisclosed	10	12.2K	1

The breakdowns below round their figures, except the open datasets' hours; STATS.md has every figure in full.

Who transcribed what

The combined config interleaves several transcription systems within each shard, so this is the breakdown across the whole dataset. The systems differ substantially, so treat them as separate sources when training.

Each row carries the system that produced it in its provider and model columns; the labels below are withheld aliases for the same systems, in the same order of size.

vendor	hours	chunks	share of hours
vendor-a	17.4K	20.5M	84.7%
vendor-b	3K	3.6M	14.4%
vendor-c	162	187K	0.8%
vendor-d	28.4	31.7K	0.1%
combined	5.5	6.5K	<0.1%

Open Indian-language datasets

Each open dataset is a config of its own, and every row names its source in source_id (the config) and source_db (the original repo or URL).

No clip is published twice. Across every open dataset, a row is left out if an earlier row already carries the same audio (identical 16 kHz samples), the same chunk_id, or the same broadcast clip, so datasets that repeat themselves (a bulletin ingested twice, a segment copied into the next file) keep one copy. The same speech recorded or encoded differently is not caught.

Hours here are exact. Held out is the part tagged as an evaluation split (see below); duplicates removed counts rows left out because another row already carried them.

config (source_id)	dataset	transcripts	languages	hours	of which held out	chunks	shards	duplicates removed
yodas3	YODAS3	youtube asr captions	10: Hindi, Punjabi, Bengali …	66,650.61	0.00	22,725,807	7,387	46,082
newsonair	NewsOnAIR preserved audio	website scripts where verified; otherwise untranscribed	5: Hindi, English, Telugu …	12,228.33	0.00	1,572,952	1,791	176
indicvoices	IndicVoices	human	22: Hindi, Tamil, Assamese …	11,855.81	125.96	6,222,099	1,346	24
nptel	NPTEL	human	10: English, Tamil, Hindi …	7,261.47	0.00	3,073,312	754	3,325
shrutilipi	Shrutilipi	aligned bulletin text	16: Hindi, Marathi, Tamil …	3,926.75	0.00	2,206,066	480	19,731
chaashini	Chaashini	asr	33: Hindi, English, Urdu …	2,921.85	0.00	1,376,551	403	15,435
worldspeech	WorldSpeech	asr	16: Hindi, Tamil, Marathi …	1,953.75	88.81	510,473	215	2,793
vaani	Vaani	human	61: Hindi, Telugu, Bengali …	1,928.26	386.61	1,317,716	214	2,523
pretraining-v1-cv22-sidon	Common Voice 22, restored (Pretraining-V1)	read prompts	13: Bengali, Tamil, Urdu …	1,804.21	0.00	1,425,647	230	12,519
pretraining-v1-kathbath	Kathbath (Pretraining-V1)	read prompts	12: Tamil, Marathi, Kannada …	1,475.20	0.00	805,721	207	0
rasa	Rasa	read prompts	22: Bengali, Punjabi, Assamese …	1,129.36	113.65	634,166	128	2,506
pretraining-v1-syspin	SYSPIN (Pretraining-V1)	read prompts	9: Magahi, Chhattisgarhi, Bengali …	1,080.52	12.18	511,224	151	275,401
pretraining-v1-indictts	IndicTTS (Pretraining-V1)	read prompts	13: English, Gujarati, Assamese …	502.07	0.00	282,243	64	11,765
uwc-hindi	ASR-UWC Hindi ASR corpus	not documented	Hindi	429.42	0.00	266,134	47	23,175
spring-inx	SPRING-INX Hindi	human	Hindi	234.69	47.32	91,926	33	114
bhojpuri-rural	Rural Women Bhojpuri	human	Bhojpuri, Hindi	202.69	1.26	78,805	23	0
fleurs	FLEURS	read prompts	14: Marathi, Sindhi, Malayalam …	177.53	53.29	51,938	20	0
pretraining-v1-msft-indian	Microsoft Indian-language speech (Pretraining-V1)	read prompts	Gujarati, Tamil, Telugu	134.92	0.00	115,392	18	0
pretraining-v1-spicor	SPICOR Indian English (Pretraining-V1)	read prompts	English	99.34	1.12	50,423	15	0
svq	Simple Voice Questions	read prompts	10: Bengali, Telugu, Urdu …	82.72	82.72	53,375	9	0
all-hindi-asr	All Hindi ASR v1.1 (filtered)	not documented	Hindi	66.38	0.00	27,393	9	14,246
pretraining-v1-synthetic-v1	synthetic_v1 (Pretraining-V1)	synthetic	9: Hindi, Bengali, Kannada …	21.04	0.00	9,941	3	0
lahaja	Lahaja	human	Hindi	12.32	12.32	6,152	2	0
navana-hindi	Navana Hindi ASR Benchmark	human	Hindi	11.16	11.16	7,649	2	0
pretraining-v1-ivr	IndicVoices-R (Pretraining-V1)	human	22: Dogri, Maithili, Telugu …	6.49	0.00	2,094	8	653,567
pretraining-v1-elise-hindi	Elise Hindi (Pretraining-V1)	not documented	Hindi	2.36	0.00	1,147	1	0
pretraining-v1-multilingual-tts	multilingual_tts (Pretraining-V1)	synthetic	Bengali, Hindi, Urdu	1.95	0.00	991	2	0
all open datasets				116,201.20	936.40	43,427,337	13,562	1,083,382

• yodas3: Whole YouTube videos in 10 Indian languages, cut into utterances on the words of their auto-generated (ASR) captions: noisy, machine-made labels. Uploader subtitle tracks (translations, lyrics) are not what was said and are left out; so are chunks not mostly in the language's own script.
• newsonair: 50,878 rows (135.4 h) carry All India Radio's own bulletin scripts, matched to the recording and checked against the audio word by word; the recording is cut at its pauses around them. The other rows are untranscribed: exclude them from supervised ASR. Alignment is automated, not human-reviewed. Original recordings and scripts remain archived.
• indicvoices: Crowd-recorded read, extempore and conversational speech from across India, transcribed and verified by AI4Bharat. The text is the verbatim transcript (what was said, not the standardised text); rows marking part of the audio unintelligible are left out. Its valid split (the IndicVoices benchmark) is tagged valid.
• nptel: NPTEL university lectures from AI4Bharat's BhasaAnuvaad: the original lectures in Indian English (en) and voice-over dubs in 9 Indian languages, whose text is the translated script the artist read. Kept only where the text matches an ASR pass (alignment >= 0.8, stored in confidence) and the clip is 0.5-30 s. Numbers are written as digits.
• shrutilipi: All India Radio news read against published bulletin text, aligned automatically.
• chaashini: Kapture CX's quality-gated corpus of clean single-speaker speech in 33 Indian languages and Indian English: clips cut at pauses from public spoken-word YouTube recordings (talks, interviews, narration, lectures, podcasts). The filtering is Chaashini's: recordings dominated by music or singing dropped, speech activity detection, one speaker per clip (diarised, overlaps removed), DNSMOS, SNR and music/noise scores within strict limits (borderline clips enhanced and re-scored, flagged enhanced), and ASR transcripts kept only with plausible character rates and confident recognition, so the text is that ASR's output. Rows keep Chaashini's clip id, source recording, speaker and segment index; confidence is its ASR confidence, language_confidence and speech_ratio are its own, and its other scores are in vad_detail under chaashini, so its filters still apply (e.g. dnsmos_ovrl >= 3.2 and enhanced == false). Its source videos whose transcripts are largely found in the yodas3 config (compared by content: yodas3's video ids are anonymised) are left out as duplicates. Cite: Kapture CX, "Chaashini: a quality-gated multilingual Indian speech corpus" (2026).
• worldspeech: Every Indian-language config: All India Radio news in 16 languages, Hindi legislatures and Mann Ki Baat, and the Kerala, Maharashtra and Punjab assemblies, plus Bengali (Bangladesh), Nepali (Nepal), Tamil (Sri Lanka) and Urdu (Pakistan); the config is in config_id. The text is WorldSpeech's asr_transcript (machine output, within CER 0.3 of the official record); confidence is 1 - CER.
• vaani: The transcribed part of IISc/ARTPARK's Vaani: people in 165 districts describing an image, in 64 languages and dialects. Annotation markup is stripped the way the Vaani benchmark scores, and rows marking audio unintelligible are left out. Its validation and test splits keep those tags; every segment of a recording Vaani-Benchmark-V1.0 took segments from is tagged benchmark.
• pretraining-v1-cv22-sidon: Common Voice 22 clips cleaned by the Sidon speech-restoration model. Only the rows in Indian languages are taken (13, most of them Bengali, Tamil and Urdu); its English, Pashto, Dhivehi and Saraiki are not.
• pretraining-v1-kathbath: AI4Bharat's read speech in 12 Indian languages from speakers across India. Its own valid/test splits are not marked in this copy, so every row is tagged train.
• rasa: Expressive studio speech recorded for TTS: the text is the prompt the speaker read.
• pretraining-v1-syspin: IISc SYSPIN TTS recordings: studio speech of a few speakers each in Hindi, Bengali, Marathi, Kannada, Telugu, Maithili, Magahi, Chhattisgarhi and Bhojpuri; rows named EVAL keep that tag.
• pretraining-v1-indictts: IIT Madras TTS recordings: studio speech of a few speakers each in 13 Indian languages, and English read by Indian speakers.
• uwc-hindi: Audio with transcripts, labelled Hindi by its publisher, which does not document where either comes from. Most is Hindi, but some files hold Bhojpuri, Maithili and Marathi as well, tagged hi here like the rest.
• spring-inx: Hindi phone conversations and monologues from IIT Madras (NLTM), R1 and R2 'clean' releases, all splits. R2's re-releases of R1 audio and R1's duplicate recordings are taken once. Every eval utterance is tagged eval: all of an eval recording (under any of its names) and every utterance of the original R1 and R2 test sets, which the clean releases moved into train/dev.
• bhojpuri-rural: Bhojpuri (and some Hindi) from rural women. Only train_real (400 clips) and benchmark (444) are real recordings; train_synthetic, about 99% of the rows, is synthetic speech.
• fleurs: Read FLoRes sentences in 14 Indian-language configs (as, bn, gu, hi, kn, ml, mr, ne, or, pa, sd, ta, te, ur), up to three speakers per sentence (config_id names the sentence). Its validation and test splits keep those tags.
• pretraining-v1-msft-indian: Read Gujarati, Tamil and Telugu sentences, about 45 hours of each.
• pretraining-v1-spicor: English read by Indian speakers (IISc SPICOR project); rows named EVAL keep that tag.
• svq: Short spoken questions recorded on phones in 12 Indian locales (including Indian English and Bangladeshi Bengali and Pakistani Urdu), each in one of four real conditions: clean, background speech, media, traffic. Published as an evaluation collection, so every row is tagged test.
• all-hindi-asr: Hindi audio with transcripts; the dataset does not document where either comes from.
• pretraining-v1-synthetic-v1: Synthetic speech (TTS) of conversational lines in eight Indian languages; event tags such as are removed from the text. It labels some Marathi lines Hindi; those are tagged mr (the published label stays in config_id).
• lahaja: Hindi from 132 speakers in 83 districts, read and extempore, published as a test set only (tagged test).
• navana-hindi: Hindi test clips drawn from CommonVoice, IndicTTS, Kathbath and MUCS (named in config_id) with their human references, tagged test.
• pretraining-v1-ivr: IndicVoices clips enhanced for TTS (48 kHz). The indicvoices config already carries about 98% of them (left out here as duplicates). IndicVoices-R deletes the marks from the transcripts, so clips whose IndicVoices transcript has one are left out as the indicvoices reader leaves out their originals. What remains is clips missing from IndicVoices' release.
• pretraining-v1-elise-hindi: One female voice reading Hindi lines translated from the English Elise set; likely synthetic.
• pretraining-v1-multilingual-tts: Synthetic code-switched speech (TTS); only rows in Indian languages, alone or with English, are taken (about a tenth of it).

NewsOnAIR: same format, transcript availability is explicit

newsonair uses the same Parquet columns and embedded 16 kHz mono FLAC as every other config. Audio awaiting verified transcripts has source_collection=untranscribed, empty text, and null is_unintelligible. Missing transcripts do not mean silence or unintelligible speech. Verified website-script rows use source_collection=website_script_aligned and provider=website_script. A recording whose bulletin script was matched to it and checked against the audio is re-cut at its pauses: each verified stretch becomes a row carrying the script's own words, and the rest of the recording stays untranscribed rows, so no audio is lost or repeated. Script provenance is in the first word's transcript_source object in words; word times are milliseconds from the row's start. Alignment scores are not ASR confidence. Unmatched scripts/PDFs remain preserved separately; no generated ASR text is passed off as website text. Language is unlabelled until verified; no language is guessed from a station name.

# Use after loading or concatenating configs, before supervised ASR training:
paired = ds.filter(lambda r: r["source_collection"] != "untranscribed")

Every row of every open dataset is included, evaluation sets too: those rows carry their split in source_collection (valid, validation, dev, test, eval, benchmark), so they can be left out of training in one filter.

from datasets import load_dataset

ds = load_dataset("kapturecx/bolAIndia", "yodas3", split="train")
held_out = ('valid', 'validation', 'dev', 'test', 'eval', 'benchmark')
ds = ds.filter(lambda r: r["source_collection"] not in held_out)
ds = ds.filter(lambda r: r["language_code"] == "ta")                     # one language

In open-dataset rows the call-centre columns carry the dataset's own provenance:

column	open-dataset rows
provider	who wrote the transcript: human, captions, asr, read prompts ...
model	the dataset's name
source_db	the original dataset: HF repo id or URL
source_collection	its split: train, or one of the held-out names above
config_id	the dataset's own subset (language, district, config)
mongo_id / conversation_id / call_sid	its utterance, recording and speaker ids
recording_url	where the audio sits inside the original dataset

confidence holds the dataset's own quality score where it has one (NPTEL: text-to-ASR alignment; WorldSpeech: 1 - CER) and is empty otherwise; VAD and phone-hash columns are empty.

Languages

Audio per language across every source, as the transcription system or the dataset labelled it (the language_code column). unlabelled rows carry no language code.

language	code	hours	chunks	share of hours
Hindi	hi	70K	38.5M	51.2%
unlabelled	—	12.1K	1.5M	8.8%
English	en	8K	7.2M	5.9%
Punjabi	pa	7.8K	2.3M	5.7%
Bengali	bn	7.5K	3.2M	5.4%
Tamil	ta	5.3K	2.5M	3.9%
Telugu	te	4.4K	1.9M	3.2%
Marathi	mr	4K	1.6M	2.9%
Kannada	kn	3K	1.4M	2.2%
Malayalam	ml	2.9K	1.4M	2.2%
Gujarati	gu	1.8K	859K	1.3%
Odia	or	1.3K	703K	0.9%
Urdu	ur	1.2K	672K	0.8%
Assamese	as	1.1K	613K	0.8%
Nepali	ne	956	453K	0.7%
Maithili	mai	693	361K	0.5%
Bodo	brx	597	353K	0.4%
Sanskrit	sa	591	290K	0.4%
Manipuri	mni	572	299K	0.4%
Dogri	doi	537	232K	0.4%
Sindhi	sd	493	256K	0.4%
Santali	sat	469	256K	0.3%
Kashmiri	ks	402	223K	0.3%
Konkani	kok	354	203K	0.3%
Bhojpuri	bho	245	116K	0.2%
Chhattisgarhi	hne	138	69.3K	0.1%
Magahi	mag	128	75.7K	<0.1%
Chakma	ccp	49.3	40.7K	<0.1%
Garo	grt	44.2	39.1K	<0.1%
Rajasthani	raj	29.7	15.9K	<0.1%
Nagamese	nag	21.1	14.4K	<0.1%
Mizo	lus	18.8	11.1K	<0.1%
Wancho	nnp	11	10.4K	<0.1%
Garhwali	gbm	8.4	5.7K	<0.1%
Marwari	mwr	7.8	4.2K	<0.1%
Bajjika	vjk	4.4	2K	<0.1%
Kokborok	trp	3.8	3.1K	<0.1%
Khortha	x-khortha	3.4	2.5K	<0.1%
Angika	anp	3	1.6K	<0.1%
Tulu	tcy	2.8	1.9K	<0.1%
Kumaoni	kfy	2.3	1K	<0.1%
Halbi	hlb	2	979	<0.1%
Sumi	nsm	1.5	914	<0.1%
Idu Mishmi	clk	1	1.1K	<0.1%
Malvani	x-malvani	0.93	605	<0.1%
Sadri	sck	0.9	660	<0.1%
Bhili	bhb	0.79	552	<0.1%
Bundeli	bns	0.7	392	<0.1%
Surgujia	sgj	0.63	525	<0.1%
Karbi	mjw	0.62	596	<0.1%
Rengma	nre	0.47	352	<0.1%
Ao	njo	0.47	281	<0.1%
Gondi	gon	0.41	268	<0.1%
Haryanvi	bgc	0.39	277	<0.1%
Chakhesang	x-chakhesang	0.35	269	<0.1%
Rongmei	nbu	0.27	190	<0.1%
Awadhi	awa	0.25	172	<0.1%
Kurukh	kru	0.22	131	<0.1%
Surjapuri	sjp	0.21	177	<0.1%
Sambalpuri	spv	0.2	149	<0.1%
Dhundari (Jaipuri)	dhd	0.17	81	<0.1%
Beary	x-beary	0.16	120	<0.1%
Angami	njm	0.16	104	<0.1%
Tagin	tgj	0.12	93	<0.1%
Nyishi	njz	0.11	77	<0.1%
Thethi	x-thethi	0.11	80	<0.1%

NewsOnAIR original archive (separate from training rows)

Full original audio, original scripts/documents, and catalogue snapshots are preserved under archives/newsonair/. Collection is incremental, not complete. These are not aligned audio/transcript pairs and are excluded from all Parquet hours, chunks, language tables, and dataset configs above.

Any standardized newsonair chunks are counted separately in the Parquet tables. They derive from these originals: do not add archive hours to Parquet hours.

Verified archive checkpoint: 2026-10-05T04:49:38.163674+00:00. Statistics refresh periodically; newer archive commits may not yet be reflected in this snapshot.

preserved originals	count
unique audio hours	12,213.53
unique audio files	80,784
unique script/document files	25,990
unique HTML script pages	3,873
unique catalogue snapshots	10,636
all unique objects	121,283
object bytes (excluding manifests)	364,744,671,267
currently published asset URLs (not unique files)	110,807
verified publication manifests	9,081

Objects and audio duration are counted once per SHA-256, including retained versions of changed source URLs. URL aliases do not add another stored copy. Different encodings or overlapping recordings are not deduplicated by this measure. Scripts become transcripts only where their text verifies against the audio (see newsonair above); original rights are not asserted to be an open licence.

Browse originals · Machine-readable archive statistics

Loading

Rows are keyed by chunk_id. In the call-centre configs a chunk that two vendors transcribed differently has one row per vendor (same chunk_id, different provider and text). To keep one transcript per clip, keep the first row of each chunk_id:

seen = set()
ds = ds.filter(lambda r: not (r["chunk_id"] in seen or seen.add(r["chunk_id"])))

from datasets import load_dataset

ds = load_dataset("kapturecx/bolAIndia", "combined", split="train")
print(ds[0]["text"], ds[0]["confidence"])

Fields

field	meaning
audio	16 kHz mono chunk of the human channel
text	provider transcript; empty means nothing intelligible was heard
confidence	provider confidence for the chunk (0-1)
word_confidence_mean / word_confidence_min	aggregated per-word confidence
words	JSON per-word timings and confidences
language_code / language_confidence	detected language and its confidence
is_unintelligible	true when the provider returned no text
provider / model / source_id	which system produced the transcript
speech_ratio / vad_detail	fused and per-engine VAD speech ratios
channel	side of the dual-channel recording (human)
source_db / mongo_id / conversation_id	provenance of the call; resolving these needs access to the source database
recording_url / s3_key / bucket	blank by design — the source recording is not distributed with this dataset
chunk_start_ms / chunk_end_ms	position of the chunk inside that recording

Phone numbers are stored only as salted hashes.