One model to transcribe, translate, and navigate 22 Indian languages

Sarvam AI is updating its speech recognition stack with Saaras V4, combining an audio encoder with a 3B-parameter hybrid language model. It supports five transcription formats while targeting Indian languages, English accents, code-mixed speech, and challenging real-world audio.

Five representations of the same recording

A conversation mixing Hindi and English can be captured exactly as spoken, normalized, romanized, or directly translated. Saaras V4 handles these operations within a single model instead of relying on separate post-processing stages.

Five modes are available. `verbatim` preserves every spoken word in its native script, `transcribe` normalizes elements such as numbers and dates, `codemix` keeps English words in Latin script within native-language text, `translit` romanizes the entire transcript, and `translate` produces an English translation.

Sarvam AI is targeting a common characteristic of multilingual speech in India: switching between languages within the same conversation or sentence. A 3B-parameter decoder trained in-house

The architecture combines an audio encoder with Sarvam-3B, a 3B-parameter autoregressive hybrid state-space language model trained from scratch by Sarvam AI.

The encoder first converts the audio waveform into acoustic representations. A temporal-downsampling adapter then compresses them along the time axis before projecting them into the language model's embedding space. The decoder processes those features alongside textual instructions and generates the transcript token by token.

That temporal compression also helps longer recordings remain within the model's context budget. 4.35% average WER across seven English benchmarks

Sarvam AI reports an average 4.35% WER across seven English evaluation suites covering meetings, podcasts, audiobooks, financial calls, speeches, and multiple English accents.

In the company's published comparison, ElevenLabs Scribe v2 records 4.62%, AssemblyAI 4.66%, and Deepgram Nova 3 5.57%. Six datasets use results published through Hugging Face's Open ASR Leaderboard, with Svarah added to evaluate Indian-accented English.

The measurements follow the Open ASR Leaderboard's normalization and scoring methodology, although the combined comparison is published by Sarvam AI. Indian languages sit at the center of the evaluation

The difference becomes more pronounced in evaluations focused on Indian languages. On Vistaar, Saaras V4 records an overall LLM-WER of 11.12% in Sarvam AI's results, compared with 14.44% for Scribe v2, 23.33% for Deepgram Nova 3, and 25.89% for GPT-4o Transcribe.

Sarvam reports two metrics here. Traditional WER counts substitutions, deletions, and insertions against the reference transcript. LLM-WER adds semantic adjudication intended to distinguish actual transcription errors from orthographic or formatting differences that preserve the underlying meaning.

The model can also identify the spoken language directly from audio. Sarvam reports a 5.22% identification error rate across all 22 evaluated Indian languages, falling to 2.9% across the ten most widely spoken languages in the group. Give the model the words it is likely to miss

Saaras V4 also supports keyterm prompting, allowing applications to provide words or phrases that should receive additional attention during transcription. These can include names, products, acronyms, or domain-specific terminology.

On AI4Bharat's IndicContextEval, Sarvam reports a 16.03% WER in the L5 configuration, where domain entities are provided in native script alongside the target language.

The system is also designed for recordings affected by wind, compression, clipping, traffic, and other interference, as well as code-mixed speech and dialect variation. Under 150 ms to the first token

For real-time applications, Sarvam reports a time to first token below 150 ms in streaming mode. Multi-minute recordings can also be processed natively.

The REST API handles synchronous transcription for audio under 30 seconds, while the batch API accepts files up to two hours and can enable speaker diarization. A WebSocket interface provides partial transcripts for live voice applications.

Python and Node.js SDKs are available, alongside integrations for Vercel AI SDK, LiveKit Agents, and Pipecat Agents.