asr-hindi

Hindi ASR audio (16 kHz FLAC), one half of the BiBo voice 1000 h mix (600 h Hindi / 400 h English), built by BiBo voice/asr/build_mix.py. Bad rows were REMOVED, never relabelled: Emilia keeps rows whose two independent transcripts agree (<= 5% WER) with audio quality PQ >= 6.5; Numo keeps its own WER <= 6% and synthetic-speech score <= 0.2; NPTEL drops rows a Qwen3-ASR pass scores above 15% WER (on a Gemini audit the 25-50% band was 90% bad: labels running past the audio cut). Text keeps each source's original casing / markup; the training normalisation is in voice/asr/prep_train.py.

Configs: default 602.2 h.

source	lang	hours	rows	train use
hinglish	hi	20.0	9046	x1
indicvoices_hi	hi	300.0	163428	x1
kathbath_hi	hi	50.0	34316	x1
lahaja	hi	12.2	5667	x1
numo_hi	hi	120.0	9652	x1
vaani_hi	hi	100.0	49142	x1

Licenses: each row keeps its source's terms -- MrDragonFox/EN_Emilia_Yodas_616h CC-BY-4.0, psdn-ai/numo-indic-speech CC-BY-4.0, skbose/indian-english-nptel-v0, edinburghcstr/ami CC-BY-4.0, SALT-NLP/spotify_podcast_ASR, sawradip/phone-asr-data, facebook/voxpopuli CC0, MLCommons/peoples_speech CC-BY / CC-BY-SA, ai4bharat/Svarah, yashtiwari/PaulMooney-Medical-ASR-Data, ai4bharat/IndicVoices CC-BY-4.0, psk/vaani-asr (Vaani CC-BY-4.0), ai4bharat/Kathbath CC-BY-4.0, agarwalayushi/hinglish CC-BY-4.0, ai4bharat/Lahaja MIT. Svarah and Lahaja are public test sets: models trained on this cannot report their official numbers.