.png){kind=logo}
Saaras V4 uses a 3-billion-parameter decoder trained from scratch, giving Sarvam full control over how the model handles Indian speech.
The model produces five transcript formats from one audio file, removing the need for separate tools at each stage.
Keyterm prompting lets developers guide recognition toward brand names and technical terms, a feature aimed squarely at real business use.
A phone call in India rarely stays in one language for long. A customer might open in Hindi, switch to English mid-sentence, and close with a regional word that has no clean translation. Most speech recognition tools were not built for this kind of switching. They expect one language at a time, spoken clearly, without interruption.
Sarvam AI, an Indian startup working on language technology, has released Saaras V4 to close that gap. The model reads code-mixed speech, handles background noise, and adjusts to regional accents without extra tuning. That focus places it in a different category from ASR tools designed mainly for English.
Saaras V4 runs on two main parts: an audio encoder and a language decoder. Sarvam says the decoder, a 3-billion-parameter hybrid state-space model, was trained from the ground up rather than adapted from an existing system. That choice gives the company direct say over how the model learns Indian sounds and speech patterns, rather than inheriting the habits of a model built for other languages first.
The result shows up in the numbers. Sarvam reports state-of-the-art performance across all 22 Indian languages the model supports, including several with limited digital speech data available for training. In English, the model also posts the lowest average error rate across seven public benchmarks, covering Indian accents, international speech, meetings, and financial audio.
The process happens in three steps. First, an audio encoder converts sound into embeddings that capture how a word is pronounced. Next, an adapter compresses these embeddings along the time axis, shrinking the amount of data the decoder needs to handle. Finally, the decoder reads this compressed audio alongside any prompt text and builds the transcript one token at a time.
Sarvam reports a response time under 150 milliseconds for the first output token, which supports real-time streaming.
Long recordings move fast too: a multi-minute file gets processed within a second, according to the company. That speed matters for live captions, voice assistants, and call center tools that cannot afford a delay.
A single audio clip means different things to different teams. A compliance officer needs an exact record. A CRM system wants clean, normalized text. An analytics tool might only need an English summary. Saaras V4 handles all of it from one file, without routing audio through separate tools.
For a sentence like "Yaar, mujhe kal ka meeting reschedule karna hai," each mode produces a distinct, usable version, from the raw Hindi to a clean English translation, without a second processing step.
Generic models often stumble on brand names, technical terms, or local place names. Saaras V4 addresses this with key term prompting, a feature limited to this version. Developers can supply up to 50 specific terms, each capped at 64 characters, and the model gives those words added weight during recognition.
This does not force the term into every transcript. It simply raises the odds that a name like a product brand or a technical acronym gets picked up correctly, instead of being guessed at from context alone.
A support platform, for instance, could load in product names and branch locations ahead of a call, cutting down on the kind of misheard terms that frustrate both customers and analysts. Sarvam's work on this feature was presented at Interspeech 2026.
Saaras V4 is not simply another entry in the ASR race. Sarvam is positioning it as core infrastructure for voice-based products across India, from customer support bots to meeting tools built for teams that switch languages without thinking about it.
Building the decoder from scratch gives the company a base it can keep shaping around Indian speech, rather than working around the limits of a model designed elsewhere first.
For an industry chasing English benchmarks, Saaras V4 makes a different bet: that the next real advance in speech AI depends on solving for how India actually speaks.
Also Read: India’s AI Breakthrough: SraVaani Brings Speech Recognition to 65 Indian Languages
The next phase of voice AI in India will likely be judged less on raw accuracy scores and more on how naturally these systems handle real conversation, code-switching, background noise, and regional variation included. Models built around that reality, rather than retrofitted for it, stand to shape how voice interfaces evolve across the country's next wave of digital users.
1. What is Sarvam AI Saaras V4?
Saaras V4 is Sarvam AI’s automatic speech recognition model designed to process multilingual and code-mixed speech across 22 Indian languages and English.
2. How does Saaras V4 handle code-mixed speech?
The model is designed to recognize speech that switches between languages within the same conversation, such as Hindi-English speech commonly used in India.
3. What output formats does Saaras V4 support?
Saaras V4 supports verbatim transcription, normalized transcription, code-mixed output, transliteration and translation.
4. What is keyterm prompting in Saaras V4?
Keyterm prompting allows developers to provide domain-specific terms such as product names, brands, places and technical vocabulary to improve recognition of important words.
5. What makes Saaras V4 relevant to Indian voice AI?
Saaras V4 is designed around Indian speech conditions, including multiple languages, code-switching, dialect variation and noisy audio, making it relevant to voice assistants, customer support and transcription applications.
