Sarvam AI Releases Saaras V4 for 22 Indian Languages and Global English

Sarvam has not released the model weights, and its SageMaker self-hosting documentation currently covers only Saaras v3. The default model remains Saaras v3, but V4 uses the same request shape, so switching is a one-line change, Sarvam says.

Saaras V4 is an encoder-decoder system. An audio encoder converts the waveform into embeddings that carry phonetic and acoustic detail. A temporal-downsampling adapter then shortens that sequence and projects it into the language model's embedding space, keeping long recordings inside the decoder's context budget. The decoder is Sarvam-3B, a 3-billion-parameter hybrid state-space language model trained from scratch in-house. It reads the audio features alongside a text prompt and emits the transcript autoregressively, feeding each token back as input for the next.

Sarvam evaluated English on seven datasets. Six come from Hugging Face's Open ASR Leaderboard: AMI, GigaSpeech, LibriSpeech clean, LibriSpeech other, SPGISpeech and VoxPopuli. The seventh is AI4Bharat's Indian-accented Svarah. Scoring follows the leaderboard's normalization code. Saaras V4 posts the lowest average word error rate among the models Sarvam benchmarked.

On the Indic side, Sarvam reports results on Vistaar across 10 Indian languages using WER and LLM-WER. LLM-WER adds a semantic check that separates real meaning errors from harmless spelling or formatting variants common in Indic scripts. On Kathbath Noisy, measured with LLM-WER, Sarvam says Saaras V4's error rate is under half that of Deepgram Nova-3 and GPT-4o Transcribe. That set includes compressed, clipped and background-heavy recordings. On verified IndicVoices utterances, language identification error is 2.9% across the top 10 Indian languages and 5.22% across all 22. All these numbers are vendor-reported. Independent reproduction has not been published yet.

The same audio can return five representations, selected through the mode parameter: transcribe, the default, returns native script with numbers and dates normalized; verbatim keeps every word as spoken, including fillers and spoken numbers; codemix returns native script with English words left in English; translit returns the full utterance in Latin script; and translate returns an English translation with numbers normalized. Sarvam says handling these inside the model removes post-processing steps that can compound errors.

Keyterm prompting is new in V4 and works only with saaras:v4. Users pass a JSON list under keyterms, with up to 50 terms of 64 characters each. Keyterms bias recognition but do not guarantee output. Sarvam says codemix mode should be used when a brand such as PhonePe must stay in Latin script. On IndicContextEval, a paper presented at Interspeech 2026, Sarvam reports 16.03% WER in the L5 keyword-prompting setting and says that is the lowest score on the benchmark.

Sarvam says streaming works over WebSocket with partial results and time to first token below 150 milliseconds. REST offers synchronous transcription for clips up to 30 seconds. Batch jobs are asynchronous and can process files up to two hours each, with optional speaker diarization. SDKs are available for Python 3.9+ and Node.js 18+, plus integrations for LiveKit Agents, Pipecat and Vercel AI SDK. Sarvam lists speech-to-text at ₹30 per hour for real-time, streaming and batch, and ₹45 per hour with diarization.

Sarvam benchmarked Saaras V4 against Deepgram Nova-3, ElevenLabs Scribe v2 and OpenAI GPT-4o Transcribe, using figures from each vendor's public documentation and pricing pages checked on September 26, 2026. Sarvam lists 22 Indian scheduled languages for Saaras V4, compared with 11 for Deepgram Nova-3 and 14 for ElevenLabs Scribe v2; OpenAI does not list Indian languages separately. Total language counts are 23 for Sarvam, 45+ for Deepgram, 90+ for ElevenLabs, and multilingual for OpenAI. Keyterm biasing supports up to 50 terms on Sarvam, is a paid add-on on Deepgram, supports up to 1,000 terms in batch and 50 in real time on ElevenLabs as a paid add-on, and uses a free-text prompt on OpenAI. Built-in output modes are five on Sarvam, transcript plus Smart Formatting on Deepgram, verbatim or no_verbatim on ElevenLabs, and transcript on OpenAI. Sarvam claims WebSocket real-time streaming with under 150 ms time to first token; Deepgram offers WebSocket streaming; ElevenLabs Scribe v2 Realtime is about 150 ms; and OpenAI offers file streaming with live audio through its Realtime API. Sarvam offers optional speaker diarization in its batch API. List prices are ₹30 per hour for Sarvam, $0.0052 per minute for Deepgram multilingual pre-recorded, $0.22 per hour for ElevenLabs batch, and about $0.006 per minute for OpenAI. For self-hosting, Sarvam says V4 is not available yet, though v3 is on SageMaker; Deepgram offers self-hosting, while ElevenLabs and OpenAI are cloud APIs.