Sarvam AI 发布 Saaras V4 语音识别模型

Home; /; Blog

The Next Leap in ASR for a Multilingual World

We are releasing Saaras V4, the latest generation of our speech recognition model.

This post covers the work behind Saaras V4, including the data used to train it, our pre-training approach, the architectural and decoding changes that support multiple output modes, and the evaluation methodology behind our results. We also report benchmark performance across Indian languages, English, and challenging real-world speech conditions.

Highlights

Saaras V4 uses an audio encoder, and an LLM decoder. The LLM decoder is a 3B hybrid state space model completely trained from scratch in-house; Saaras V4 achieved SOTA performance on all 22 Indian languages and reaffirms our commitment to even the most low-resource Indian languages; Saaras V4 achieves lowest WER on seven global English datasets, six being foreign English dialects and one being Indian English; You can choose the format that works best for you: Saaras V4 supports five transcript formats: transcribe, translate, verbatim, translit, and codemix; Saaras V4 is robust to noisy audios and code mixing. It also handles dialectical variation seamlessly; Its language identification is also best-in-class: a 2.9% error rate across the top ten Indian languages, and 5.22% across all 22

Inside Saaras V4

Saaras V4 pairs an audio encoder with a 3B hybrid state-space language model that we trained from scratch in-house.

We designed the system to handle the range of speech we see in practice, including code-mixing, dialectal variation, and noisy audio, while supporting five different ways of representing the same speech: verbatim transcription, normalized text, code-mixed text, transliteration, and translation.

Performance across English benchmarks

We evaluated Saaras V4 across seven English speech recognition benchmarks covering Indian English, international accents, meetings, media, finance, and other real-world speech settings. Across these datasets, Saaras V4 achieves the lowest average word error rate.

Meeting rooms, far-field microphones

Podcast, audiobook, web video

Read speech, studio conditions

Read speech, harder acoustics

Earning calls, financial vocabulary

European Parliament speeches

Indian-accented English

Evaluation methodology

Six international: taken as published at HuggingFace's Open ASR Leaderboard; One Indian accented English: Svarah published by AI4Bharat; Normalization and scoring follow the Open ASR Leaderboard's code

Indic Benchmarks for Saaras V4:

Vistaar

We evaluate Saaras V4 on Vistaar across ten Indian languages and report both standard Word Error Rate (WER) and LLM-WER.

WER remains the standard deterministic measure for speech recognition, counting substitutions, deletions, and insertions between the reference and predicted transcript. In Indic-language settings, however, some of these differences can be orthographic or formatting variations that do not materially change the spoken content.

To account for this, we also report LLM-WER, which adds a semantic adjudication step to determine whether a mismatch changes the meaning of the transcript or simply represents the same content differently.

The two metrics are therefore complementary. WER provides a reproducible baseline, while LLM-WER gives a more faithful view of errors that affect the underlying speech content.

Detailed breakdown of various benchmarks:

Benchmark WER by dataset

Contains crowd-sourced read speech from the Mozilla Foundation's Commonvoice.

Audio examples

Five Modes, One Model

Different applications need different representations of the same speech. A compliance archive may need a verbatim transcript, while a CRM workflow may require normalized text and an analytics system may need an English translation. These transformations are often handled through separate post-processing steps, which add complexity and can introduce additional errors.

Saaras V4 handles these requirements within the model itself. From the same audio input, it can produce five transcript modes: verbatim, normalized, code-mixed, transliterated, and translated.

Verbatim: The exact transcription in the native script of the spoken language; Transcribe: The exact transcription in the native script of the spoken language with numbers and dates normalised; Codemix: Transcribe mode plus keeps native english words in english to improve legibility; Translit: Does transcription but in english script. Colloquially referred to as “WhatsApp language”; Translate: The English translation of the audio with numbers normalized

Because all five come from one model, there are no extra pre-processing steps that could lead to cascading of errors.

Beyond the transcription

a. Best in class Language Identification

Saaras V4 can identify the spoken language directly from the audio and transcribe it in the corresponding native script, without requiring the language to be specified in advance. On verified IndicVoices utterances, language identification error is 5.22% across all 22 Indian languages, falling to 2.9% across the ten most widely spoken languages.

b. Keyterm prompting

Keyterm prompting lets you bias transcription toward words or phrases that are easy to miss, such as product names, people’s names, acronyms, or domain-specific terminology.

Validated as best on ai4bharat/IndicContextEval (Interspeech 2026): Saaras V4 reaches the lowest WER of 16.03% in the L5 (keyword prompting) setting, where the domain-entity list is provided in native script along with the language. For dataset details and competitor scores, see the paper.

ઢૂન માટે આંખનું કામ કરે છે અને દરેક વખતે

ડ્રોન માટે આંખનું કામ કરે છે અને દરેક વખતે

Reference keywords provided (24)

c. Built For Real-World Audio

Field audio is compressed, clipped, and full of interference. These are conditions where systems trained on clean benchmarks often fall apart. Saaras V4 is designed to remain reliable under these conditions, including noisy audio, code-mixing, and dialectal variation.

d. Low-latency streaming

For real-time voice applications, latency is as important as transcription quality. Saaras V4 supports streaming with time to first token below 150 ms, allowing downstream systems to begin responding with minimal delay and enabling more natural turn-taking.

It also handles long-form audio natively, with multi-minute recordings processed within a second.

Integrating Saaras V4 into your platform

Integrate Saaras v4 into your application with a simple API key from the Sarvam AI dashboard

Choose Your Integration Path:

REST API: Synchronous transcription for audio under 30 seconds; Batch API: Asynchronous processing for files up to 2 hours. Speaker Diarization can be enabled in batch mode; WebSocket (Real-time): Streaming transcription with partial results for live applications

SDKs & Frameworks:

Ready-to-use SDKs for Python 3.9+ and Node.js 18+.

Framework integrations available for Vercel AI SDK, LiveKit Agents and Pipecat Agents

Output Modes:

Configure transcription output: transcribe (default), translate (to English), translit (romanized), or verbatim or codemix

Quick Start:

See the Developer Quickstart for code examples and authentication details

One model across languages

With Saaras V4, you no longer have to trade off English performance against Indic language coverage. It delivers strong performance across both, while supporting low-latency streaming and reliable production workloads. This makes it possible to build multilingual voice applications with one ASR model.

Building voice agents?

End-to-end pricing with the models included, as low as ₹3 a minute.

Curious what else we're building? Explore our APIs and start creating.

Curious what else we're building?

Explore our APIs and start creating.