SraVaani 1.0
Our flagship. A 430M-parameter FastConformer with a hybrid TDT–CTC decoder, pretrained from scratch on Vaani and fine-tuned across 65 Indian languages and dialects.
Project Vaani began as a dataset of 31,255 hours of natural speech from 165 districts. That corpus now powers 20 open source speech models: multilingual and monolingual speech recognition, speech synthesis, and spoken language identification. Every one of them is released on Hugging Face under the MIT licence, free to download, fine-tune and deploy.
Flagship model
A 430M-parameter FastConformer that recognises speech across 65 Indian languages and dialects, including Tulu, Wancho, Kokborok and Bearybashe, which no commercial ASR system covers at all. Pretrained from scratch on Vaani, then grounded against 11 million audio–image pairs.
Interactive
Inference runs on Hugging Face Spaces, not on this server. The demos below are the official ARTPARK-IISc Spaces embedded in the page. Speak or upload a clip to transcribe it, or type text to hear it synthesised.
Model catalogue
Start with the task: recognise speech, synthesise it, or identify the language being spoken. Within speech recognition, choose a multilingual model if you do not know the input language in advance, or a monolingual checkpoint if you do and want something smaller and sharper.
One model, many languages. Handles code-mixed and unknown-language input without needing to be told what is being spoken. That makes it the right default for most applications.
Our flagship. A 430M-parameter FastConformer with a hybrid TDT–CTC decoder, pretrained from scratch on Vaani and fine-tuned across 65 Indian languages and dialects.
The streaming sibling: a lighter, low-latency model that emits the transcript while the person is still speaking. Built for voice interfaces, live captioning and call-centre use.
Single-language FastConformer checkpoints trained on Vaani. Smaller and faster than the multilingual model, and the right pick when you already know the input language and want to minimise compute.
OpenAI Whisper checkpoints fine-tuned on Vaani data, released across the full size ladder from tiny (37.8M) through large-v3 (2B). Pick a size to match your latency and hardware budget: tiny and small run on CPU or edge devices, large-v3 gives the best accuracy on a GPU.
Turning text back into natural Indian-language speech: the other half of a voice interface.
A compact ~0.1B parameter zero-shot TTS model covering 27 Indian languages, trained on curated Vaani audio. Use a built-in female or male voice, or clone a new one from a reference clip. Aimed at accessibility tooling, IVR systems and read-aloud content in languages mainstream TTS has never served.
The routing layer of a multilingual pipeline. Work out what is being spoken before deciding which model should transcribe it.
Identifies the spoken language directly from raw audio across 42 Indian languages, a scale of coverage that opens up genuinely new territory for Indic speech pipelines. A 0.8B-parameter feature-extraction model.
Complete index
Filter by task or by language coverage.
| Model | Task | Coverage | Family | Params | Languages | Details |
|---|---|---|---|---|---|---|
| SraVaani-1.0 | ASR | Multilingual | FastConformer TDT–CTC | 430M | 65 | Model page → |
| SraVaani-0.5-live | ASR | Multilingual | Streaming | – | Multi | Open ↗ |
| DhVaani-0.5 | TTS | Multilingual | Zero-shot TTS | 0.1B | 27 | Open ↗ |
| Vaani-LID_v0 | LID | Multilingual | Feature extraction | 0.8B | 42 | Open ↗ |
| Vaani-FastConformer-Hindi | ASR | Monolingual | FastConformer | – | Hindi | Open ↗ |
| Vaani-FastConformer-Kannada | ASR | Monolingual | FastConformer | – | Kannada | Open ↗ |
| Vaani-FastConformer-Telugu | ASR | Monolingual | FastConformer | – | Telugu | Open ↗ |
| Vaani-FastConformer-Malayalam | ASR | Monolingual | FastConformer | – | Malayalam | Open ↗ |
| Vaani-FastConformer-Odia | ASR | Monolingual | FastConformer | – | Odia | Open ↗ |
| whisper-large-v3-vaani-hindi | ASR | Monolingual | Whisper large-v3 | 2B | Hindi | Open ↗ |
| whisper-large-v3-vaani-telugu | ASR | Monolingual | Whisper large-v3 | 2B | Telugu | Open ↗ |
| whisper-large-v3-vaani-odia | ASR | Monolingual | Whisper large-v3 | 2B | Odia | Open ↗ |
| whisper-medium-vaani-hindi | ASR | Monolingual | Whisper medium | 0.8B | Hindi | Open ↗ |
| whisper-medium-vaani-kannada | ASR | Monolingual | Whisper medium | 0.8B | Kannada | Open ↗ |
| whisper-medium-vaani-telugu | ASR | Monolingual | Whisper medium | 0.8B | Telugu | Open ↗ |
| whisper-medium-vaani-bengali | ASR | Monolingual | Whisper medium | 0.8B | Bengali | Open ↗ |
| whisper-small-vaani-hindi | ASR | Monolingual | Whisper small | 0.2B | Hindi | Open ↗ |
| whisper-small-vaani-kannada | ASR | Monolingual | Whisper small | 0.2B | Kannada | Open ↗ |
| whisper-small-vaani-tulu | ASR | Monolingual | Whisper small | 0.2B | Tulu | Open ↗ |
| whisper-tiny-vaani-hindi | ASR | Monolingual | Whisper tiny | 37.8M | Hindi | Open ↗ |
Training data
Everything the models were trained and measured on is public, released before the models themselves were.
The full multimodal dataset: 22.6M rows of speech and images from 165 districts, released under CC-BY-4.0.
The transcribed portion used for supervised fine-tuning across the language families.
The evaluation set behind the published WER figures, with a public leaderboard on Hugging Face.
Every model on this page is open source, trained on data we open-sourced first. Download the weights, fine-tune them, ship them commercially, and tell us what you build.