Loading…
The Indian Institute of Science, Bengaluru released SraVaani-1.0, an open-source automatic speech recognition model covering 65 Indian languages and dialects.
It was built by the SPIRE Lab and ARTPARK at IISc with support from Google, and is hosted on Hugging Face under the permissive MIT licence.
The 65 languages include the 22 tongues listed in the Eighth Schedule of the Constitution plus 43 regional and tribal varieties such as Garo, Kokborok, Tulu and Bundeli.
Training used Project Vaani, a speech corpus of 31,255 hours recorded from 156,534 speakers across 165 districts in 28 states and 3 union territories.
For Garo, a language with almost no existing speech technology, SraVaani reported a 9.5% word error rate against 69.4% for the next-best system tested.
| Parameter | SraVaani-1.0 (IISc / ARTPARK, 2026) | IndicVoices and IndicASR (AI4Bharat-IIT Madras with Sarvam AI, 2024) |
|---|---|---|
| Languages covered | 65 Indian languages and dialects | 22 Eighth Schedule languages |
| Speech data | 31,255 hours unlabelled plus about 31,263 hours labelled | 7,348 hours collected, of which 1,639 hours transcribed |
| Speakers and districts | 156,534 speakers, 165 districts | 16,237 speakers, 145 districts |
| Institutional home | IISc Bengaluru (SPIRE Lab and ARTPARK) | IIT Madras (AI4Bharat), with Sarvam AI |
WER is the standard yardstick for judging a speech recognition system. It counts the words the machine substituted, deleted or inserted wrongly, and expresses that total as a percentage of the words actually spoken. Lower is better: a 9.5% WER means about one word in ten was mistranscribed. Because insertions are counted, WER can exceed 100% when a system produces far more words than were spoken, which is what happens when a model is fed a language it has never learned. This is why SraVaani's Garo figure is the headline claim of the release: 9.5% against 69.4% for the next-best system tested is not an incremental gain but the difference between a usable tool and an unusable one.
Simple Analogy: Think of a dictation clerk. If you dictate 100 words and the clerk gets 10 wrong, misses some and invents others, the error rate is the count of all those slips out of 100. A clerk who invents more words than you spoke can score worse than 100%.
Lists 22 scheduled languages; SraVaani covers all 22 and adds 43 non-scheduled varieties, making it a ready example for questions on linguistic inclusion.
The MeitY platform for Indian-language AI; Project Vaani data is open-sourced in coordination with it.
The national push to build indigenous AI compute, datasets and models; open Indic speech models are the language-data pillar of that effort.
Constitutional provisions on the development of Hindi and on protection of the distinct language and culture of minorities, the policy backdrop to language technology.
The MIT licence permits commercial reuse with attribution, unlike restrictive research-only licences, which is what makes the release usable by startups and state governments.
GS Paper 3 > Science and Technology: developments in IT and AI; GS Paper 1 > Indian Society: linguistic diversity
General Awareness > Science and Technology current affairs
General Awareness > National science and technology news
General Awareness > Current events of national importance
Technology that converts spoken audio into machine-readable text.
Percentage measure of transcription errors, counting substitutions, deletions and insertions; lower is better.
Training a model on unlabelled data by making it predict hidden parts of its own input, before any transcribed data is used.
A language with little digital text or audio data and few available language-technology tools.
A permissive open-source licence allowing free use, modification and redistribution, including commercially.