Sarvam releases Saaras V4 for multilingual speech recognition

Sarvam has released Saaras V4, a speech-to-text system that pairs an audio encoder with Sarvam-3B, a hybrid state-space language model trained from scratch. It supports English, all 22 languages in India’s Eighth Schedule, code-mixed speech, noisy recordings, automatic language identification, and five selectable output formats.

Sarvam says 10 of those Indian languages lacked another commercial automatic speech recognition service at launch. The company also reports the lowest average word error rate among the systems in its comparison across seven English benchmarks. Together, those claims position Saaras V4 for developers serving multilingual Indian users while retaining broad English support.

From waveform to token

Saaras V4 converts audio into text through three stages designed to preserve acoustic detail while compressing long recordings:

1. Audio encoding: The encoder turns the waveform into embeddings that represent phonetic and acoustic features.
2. Temporal compression: A downsampling adapter shortens the sequence along the time axis and projects it into the language model’s embedding space, reducing the context required for long audio.
3. Autoregressive decoding: Sarvam-3B reads the projected audio features alongside prompt tokens, generates one transcript token, and feeds that token back into the next decoding step.

The decoder’s hybrid state-space architecture carries a compact hidden state through the sequence while retaining selected attention mechanisms. Full attention compares tokens across an entire sequence, causing compute and memory use to rise quickly as the input grows. State-space layers keep more of that work close to linear in sequence length, which makes the design suitable for long recordings.

One model, five text forms

Each request can select one of five representations, with the model generating the chosen form directly during inference:

Direct generation reduces the need for separate normalization, translation, and transliteration services, limiting the opportunities for one stage’s errors to propagate into the next. The documented interface exposes the format as an output-mode parameter. Applications that require several representations from one recording should confirm whether the service can return multiple modes together or requires separate requests.

Claims, metrics, and caveats

Word error rate, or WER, counts substitutions, deletions, and insertions against a reference transcript, with lower scores indicating a closer literal match. Sarvam also reports LLM-WER for Indic evaluation. That metric uses a language model to judge whether a mismatch changes the meaning or reflects an equivalent spelling or orthographic form, which can prevent harmless script variations from inflating the score.

The figures in the announcement are vendor-reported, and average WER can hide large differences between languages, accents, recording conditions, and domains. LLM-WER also depends on the semantic adjudicator and evaluation prompt, making its reproducibility less straightforward than standard WER. Production comparisons should use identical audio, transcript rules, language settings, and keyword lists across providers.

Keyterm prompting lets an application supply names, brands, acronyms, and domain vocabulary before transcription begins. The decoder then assigns more probability to those terms when the audio is ambiguous. In Sarvam’s Gujarati field-audio example, adding a 24-item keyword list changes an incorrect nonsense transcription to the Gujarati equivalent of “drone.”

Three routes into production

Sarvam provides SDKs for Python 3.9 or newer and Node.js 18 or newer, along with integrations for Vercel AI SDK, LiveKit Agents, and Pipecat Agents. Switching among transcript modes requires changing the request’s output parameter.

Sarvam advertises streaming time to first token below 150 milliseconds, allowing a voice application to start processing partial text before the final transcript arrives. The company also says multi-minute recordings can be processed in under one second. End-to-end latency will additionally depend on network distance, audio buffering, endpoint detection, concurrency, and server load, so teams should measure first-token and final-transcript latency at realistic traffic levels.

Where the model fits

Saaras V4 targets voice agents, contact-center systems, field-data collection, media transcription, and analytics products serving Indian users across several languages. Its combination of automatic language detection, code-mixed transcription, native-script output, transliteration, translation, and keyterm prompting can replace several components in an existing speech pipeline.

English-only applications can compare Saaras V4 with Deepgram, ElevenLabs Scribe, GPT-4o Transcribe, and other established services using their own audio. Sarvam’s seven-benchmark result indicates competitive English recognition, while provider selection will also depend on latency, diarization, operational controls, geographic availability, and cost.

Checks before launch

A production evaluation should cover the conditions that aggregate benchmark scores tend to conceal:

• WER and semantic accuracy for every required language, accent, and script
• Code-mixed conversations with realistic language switching
• Language-identification errors on short or ambiguous utterances
• Names, product terms, acronyms, and native-script keyword prompts
• Compression, clipping, background speech, and field-recording noise
• First-token and final-transcript latency at expected concurrency
• Speaker diarization accuracy for overlapping and interrupted speech
• Normalization behavior for dates, numerals, currency, and punctuation

The release does not disclose pricing. Current rates, quotas, data-retention terms, processing regions, and security controls should be confirmed through Sarvam’s dashboard and service documentation before production deployment.