Sarvam AI releases Saaras V4: speech recognition for 22 Indian languages and global English

Sarvam AI has launched Saaras V4, the latest generation of its speech-recognition model, covering all 22 of India's scheduled languages plus English with global accents. The company claims state-of-the-art accuracy across every language. The model is live now via the API with `model="saaras:v4"`. Model weights have not been published, and the self-hosting documentation for SageMaker still references only version 3.

Under the hood sits an encoder-decoder architecture. An audio encoder converts the waveform into embeddings that capture phonetic and acoustic detail; a temporal-downsampling adapter shortens the sequence and projects it into the language model's embedding space. The decoder is Sarvam-3B, a 3-billion-parameter hybrid state-space language model built in-house from scratch. It consumes the audio features alongside a textual prompt and emits the transcript autoregressively, feeding each token back as input for the next.

Benchmarks are vendor-reported only; no independent replication has been published. On English, Sarvam evaluated 7 datasets — six from the Hugging Face Open ASR Leaderboard (AMI, GigaSpeech, LibriSpeech clean/other, SPGISpeech, VoxPopuli) plus AI4Bharat's Svarah set with Indian-accented English — and reports the lowest average word-error rate among the models tested. For Indian languages, the company measured Vistaar across 10 languages using both WER and LLM-WER, a metric that adds a semantic check to separate meaning errors from spelling or formatting quirks common in Indic scripts. On noisy audio (Kathbath Noisy) the LLM-WER is less than half that of Deepgram Nova-3 and GPT-4o Transcribe. Language ID error stands at 2.9% on the top 10 languages and 5.22% across all 22.

The same audio can be returned in five representations via the `mode` parameter: `transcribe` (default, native script with number and date normalization), `verbatim` (every word as spoken, including fillers), `codemix` (native script with English words kept in Latin script), `translit` (full Latin-script transliteration), and `translate` (English with number normalization). Sarvam argues that handling these transformations inside the model avoids post-processing stages that can compound errors. V4 also introduces keyterm prompting: pass a JSON list of up to 50 terms (64 characters each) under `keyterms` to bias recognition without guaranteeing exact output; in `codemix` mode a brand such as PhonePe remains in Latin script. On IndicContextEval (paper, Interspeech 2026) the company reports a WER of 16.03% at the L5 setting, the lowest in the benchmark.

For production, three paths are offered: WebSocket streaming with partial results and time-to-first-token under 150 milliseconds; synchronous REST for segments up to 30 seconds; and asynchronous batch for files up to two hours with optional speaker diarization. SDKs exist for Python 3.9+, Node.js 18+, plus integrations with LiveKit Agents, Pipecat, and the Vercel AI SDK. Pricing is ₹30 per hour for streaming, real-time, and batch; ₹45 per hour with diarization. Saaras v3 remains the default, and V4 uses the identical request shape so migration is a one-line change.