Dataset · ai4bharat
Speech and transcripts of 22 Indian languages
ai4bharat/IndicVoices
23.7k hours of spontaneous and read speech in 22 Indian languages, with 11.2k hours transcribed.
What's inside
The dataset contains natural and spontaneous speech: 23.7k hours of audio featuring read (8%), improvised (76%), and conversational (15%) styles. Recordings are from 51k speakers across more than 400 districts of India and cover 22 languages. As of December 23, 2025, 11.2k hours have been transcribed.
What you can do with it
- Train ASR for Indian languages
- Fine-tune a multilingual speech model
- Evaluate ASR on spontaneous speech
- Size
- 1M–10M rows
- Downloads this month
- 23.6K
- Likes
- 168
- Access
- on request on Hugging Face
Models trained on it
- theharshithh/open-sarikaAutomatic speech recognition 967
- third-intelligence/RubatoCodecAudio → audio 720
- udgamlabs/ShrutiAutomatic speech recognition 680
- Phoonnx/f5tts-multi-spring-unicodeText → speech 594
- udgamlabs/shruti-teluguAutomatic speech recognition 403
- SPRINGLab/SPRING_F5Text → speech 361
- udgamlabs/shruti-tamilAutomatic speech recognition 277
- xnpx/wav2vec2-large-xlsr-ipa-phonemesAutomatic speech recognition 271
How to load it
from datasets import load_dataset
ds = load_dataset("ai4bharat/IndicVoices")