Saaras V4 Covers All 22 Indian Languages With a 3B Hybrid Decoder

Sarvam AI has shipped Saaras V4, a speech-to-text model covering all 22 scheduled Indian languages plus global English, deployable today via API with model="saaras:v4". The release combines a 3-billion-parameter hybrid state-space decoder with five built-in output modes and a new keyterm prompting layer — a combination that puts it in direct competition with Deepgram Nova-3, ElevenLabs Scribe v2, and OpenAI GPT-4o Transcribe on Indian-language coverage, while adding structural output control those systems lack natively.

Architecture

Saaras V4 is an encoder-decoder system. An audio encoder processes the raw waveform into embeddings carrying phonetic and acoustic detail. A temporal-downsampling adapter then compresses that sequence along the time axis and projects it into the decoder's embedding space, keeping long recordings within the context budget. The decoder is Sarvam-3B — a 3B-parameter hybrid state-space language model trained from scratch in-house — which reads those compressed audio features alongside a text prompt and emits the transcript autoregressively. The model handles 8 kHz telephony and 16 kHz audio, including noisy and code-mixed speech.

Benchmarks

All figures are vendor-reported by Sarvam; independent reproduction has not been published.

On English, Sarvam evaluated seven datasets: six from Hugging Face's Open ASR Leaderboard — AMI, GigaSpeech, LibriSpeech clean, LibriSpeech other, SPGISpeech, and VoxPopuli — plus AI4Bharat's Indian-accented Svarah, scored under the leaderboard's normalization code. Sarvam reports Saaras V4 posts the lowest average WER across the models it benchmarked.

On Vistaar, covering 10 Indian languages, Sarvam reports results using both standard WER and LLM-WER, the latter adding a semantic check to separate meaning-level errors from script and formatting variants common in Indic writing systems. On Kathbath Noisy — compressed, clipped, and background-heavy recordings — Sarvam reports Saaras V4's LLM-WER is under half that of Deepgram Nova-3 and GPT-4o Transcribe.

Language identification on verified IndicVoices utterances reaches 2.9% error across the top 10 Indian languages and 5.22% across all 22.

Keyterm prompting, new to V4, accepts a JSON list of up to 50 terms at 64 characters each. On IndicContextEval L5, presented at Interspeech 2026, Sarvam reports Saaras V4 scored 16.03% WER in the keyword-prompting setting — which Sarvam describes as the lowest score on that benchmark. Keyterms bias recognition; they do not guarantee output.

Output modes and deployment

The same audio, same API call shape, returns five representations selected via the mode parameter: transcribe (default, native script with numbers and dates normalized), verbatim (every spoken word kept, including fillers and spoken-out numbers), codemix (native script with English words left in Latin script), translit (full utterance in Latin script), and translate (English translation with numbers normalized).

Streaming uses WebSocket with partial results and a vendor-reported time-to-first-token below 150 ms. REST handles clips up to 30 seconds synchronously. Batch handles asynchronous jobs up to 2 hours per file with optional speaker diarization. SDKs cover Python 3.9+ and Node.js 18+, with integrations for LiveKit Agents, Pipecat, and Vercel AI SDK. Pricing is ₹30 per hour for real-time, streaming, and batch; ₹45 per hour with diarization. Saaras v3 remains the API default; V4 uses the same request shape, so switching requires a single parameter change.

Competitive positioning

AI Mastery analysis

Training Sarvam-3B from scratch rather than fine-tuning an existing decoder is a meaningful architectural choice. Adapting a general-purpose LLM to Indic scripts typically inherits tokenization schemes optimised for Latin text, inflating sequence lengths and degrading both accuracy and inference efficiency for languages like Tamil or Malayalam. A purpose-trained hybrid state-space model sidesteps that tradeoff — though Sarvam also carries the full pretraining cost and maintenance burden, which only becomes a durable moat if the accuracy advantage holds as larger vendors deepen their own Indic coverage.

The V4 self-hosting gap is the immediate operational constraint for enterprise teams. As we have noted in our analysis of AI portability as a deployment bottleneck, model capability rarely determines where production deployments stall — data residency requirements and latency profiles do. Right now, V4 has no self-hosting path; teams with those constraints are left on v3.

Keyterm prompting warrants scrutiny proportional to its usefulness. The 16.03% WER on IndicContextEval L5 is the result of a controlled keyword-prompting evaluation; production field audio with rare named entities rarely maps cleanly onto that setting, and Sarvam is explicit that keyterms bias rather than guarantee output. Teams should plan a calibration step before relying on keyterm behaviour in production.

The broader significance is that full 22-language coverage, five deterministic output modes, and sub-₹50/hour pricing compress what previously required multiple specialist models, language-specific post-processing, and manual transliteration pipelines into a single API call. Whether Sarvam's vendor-reported accuracy claims hold under independent evaluation will determine whether V4 actually changes procurement decisions — or whether the benchmark is simply part of the product.

Sources

• Reported by: MarkTechPost: Sarvam AI Releases Saaras V4: A Speech-to-Text Model for All 22 Indian Languages and Global English