IISc Researchers Unveil AI Speech Model Covering 65 Indian Languages and Dialects

Researchers at the Indian Institute of Science (IISc), working with ARTPARK, have developed SraVaani-1.0, a multilingual speech recognition model designed to bring voice AI to Indian languages and dialects that remain largely unsupported by existing systems.

The model covers 65 Indian languages and dialects, including several low-resource and tribal languages for which the researchers say no publicly available automatic speech recognition (ASR) system currently exists.

"The vast majority of automatic speech recognition (ASR) research has focused on a handful of high-resource languages, leaving most Indic languages severely underserved by existing technology," the research paper reads.

The researchers note that systems such as IndicConformer-600M-Multilingual and Sarvam Saaras v3 primarily support the 22 languages listed in the Eighth Schedule of the Constitution.

SraVaani uses a FastConformer architecture and was trained from scratch through a three-stage pipeline. The first stage involved self-supervised pretraining on 31,255 hours of unlabelled speech from the VAANI corpus, collected across 28 states and union territories and covering 105 languages.

The second stage is a key technical feature of the model. Researchers used 11.85 million paired audio-image samples from VAANI to align speech representations with visual information. The approach is designed to help the model develop richer semantic representations without requiring additional speech transcriptions, particularly benefiting low-resource languages.

In the final stage, the model was fine-tuned on 31,263 hours of transcribed multilingual Indian speech compiled from 24 public datasets covering the 65 languages and dialects.

The researchers evaluated SraVaani against IndicConformer-600M-Multilingual, Sarvam Saaras v3 and Gemini 3 Flash across eight benchmarks, including Common Voice, FLEURS, Kathbath, RESPIN, GramVaani, MUCS and VAANI. SraVaani recorded the lowest word error rate across a large number of language-dataset combinations while remaining competitive on higher-resource languages.

One of its strongest results comes from the low-resource language segment. On Garo, for example, SraVaani achieved a 9.5% word error rate, compared with 69.4% for the next-best evaluated system, according to the researchers.

The paper also highlights that SraVaani is the only open-source-evaluated model among those compared that provides transcription for multiple low-resource and tribal Indian languages, as assessed by the VAANI benchmark.

The model is being released openly, allowing researchers and developers to build speech-based applications for communities that have historically received little support from mainstream AI systems.

The researchers argue that the implications extend beyond transcription. Speech interfaces are increasingly being used in government services, healthcare, education and accessibility tools, but limited language coverage can exclude rural and tribal communities. SraVaani is intended to address that gap by making speech recognition more inclusive across India’s linguistic landscape.

Researchers also admit that SraVaani still has limitations, including no support for Urdu and Kashmiri, unreliable language identification and limited accuracy for some low-resource languages due to insufficient training data.

It also struggles with script-aware code-switching and text normalisation. Additionally, its evaluation data comes from the same underlying dataset as training, limiting conclusions about real-world, out-of-domain performance and generalisation.