IISc Team Releases SraVaani ASR for 65 Indian Languages
Researchers at the Indian Institute of Science and ARTPARK have released SraVaani-1.0, an open-source automatic speech recognition model trained for 65 Indian languages and dialects. The 430-million-parameter system uses a three-stage pipeline built on more than 31,000 hours of speech and audio-image alignment. Its authors report the broadest evaluated Indic ASR coverage to date, including many low-resource and tribal languages.
IISc and ARTPARK researchers released SraVaani-1.0, a roughly 430-million-parameter automatic speech recognition model designed for 65 Indian languages and dialects. The research paper appeared on August 8, 2026, and the model is available through an MIT-licensed Hugging Face repository, although downloading its files requires accepting the repository's access conditions.
What the team released
SraVaani uses a FastConformer encoder with a hybrid Token-and-Duration Transducer and CTC decoder. Its authors describe it as the broadest-coverage multilingual Indic ASR model reported to date. That is a research-team claim, not an independent market ranking, but the released paper provides model architecture, training data, evaluation tables and limitations that make the comparison inspectable.
The project targets scheduled languages as well as regional, tribal and low-resource varieties. The model card lists languages including Garo, Gondi, Idu Mishmi, Kokborok, Kurukh, Tulu and Wancho alongside widely used languages such as Hindi, Bengali, Kannada, Marathi, Tamil and Telugu. Urdu and Kashmiri are not supported in this release.
A three-stage training pipeline
The paper says the team first pretrained the encoder on 31,255 hours of unlabelled VAANI speech spanning 105 languages. It then aligned speech representations with 11.85 million paired audio-image samples before fine-tuning on 31,263 hours of transcribed speech assembled from 24 public datasets across 65 languages and dialects. The Hugging Face card rounds the fine-tuning total to about 31,270 hours.
The authors evaluated SraVaani across eight public benchmarks. Where competing systems were available, they report the best accuracy on 10 of 17 languages and the lowest mean word error rate across that comparison set. They also report that 44 evaluated languages had no transcription baseline among the three comparison systems, which is the release's clearest contribution: wider measurable coverage rather than uniformly best performance everywhere.
Why the release matters
For speech engineers, SraVaani provides open weights and a documented baseline for languages that often lack public ASR tooling. The audio-image alignment stage is also a useful research direction for extracting additional supervision from the VAANI collection without requiring a transcript for every utterance.
The limitations are equally important. Training volume varies substantially by language, and the model card warns that lower-resource languages may have worse recognition accuracy. Teams evaluating deployment should therefore test the exact language, dialect, acoustic setting and domain they intend to support instead of treating the 65-language headline as a guarantee of uniform production quality.
Key Points
- 1SraVaani-1.0 is an open-source, roughly 430-million-parameter ASR model trained for 65 Indian languages and dialects.
- 2Its three-stage pipeline combines 31,255 hours of self-supervised speech pretraining, audio-image alignment and supervised multilingual fine-tuning.
- 3The authors report unusually broad low-resource coverage, but accuracy varies by language and requires deployment-specific evaluation.
Scoring Rationale
The release provides inspectable open ASR weights and benchmark coverage for many low-resource Indian languages, making it practically useful to multilingual speech researchers. Its impact is moderated by uneven per-language data and performance and by the absence of production deployment evidence.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
