Sarvam AI has unveiled Sarvam Audio, a speech recognition system that aims to move beyond plain transcription and handle complex, multilingual, multi-speaker audio common across India.
The company said the model supports 22 Indian languages and Indian English, allows apps to control how transcripts are formatted, separates overlapping speakers, and uses conversational context to improve accuracy over long calls and meetings.
The release targets a clear gap. Most automatic speech recognition systems work well on clean, single-speaker recordings. They often fail in real use cases such as crowded call centres, field workers speaking in code-mixed Hindi and English, or long customer support calls with multiple voices.
Sarvam said those failures limit the usefulness of voice interfaces in a country where speaking is often easier than typing.
Sarvam Audio is built as an extension of the company’s three-billion-parameter language model, Sarvam 3B. Instead of producing a single fixed transcript, the system lets developers choose how speech is rendered.
That includes literal word-for-word text for compliance use, normalised output with numbers and punctuation for logistics and commerce, code-mixed output that keeps English terms in Roman script, full Romanised text for chat tools, and instant translation into English for creators.
The company said that in internal tests on the IndicVoices benchmark, the company reported lower word error rates than GPT-4o-Transcribe and Gemini 3 Flash across unnormalised, normalised, and code-mixed formats. It states these gains are achieved without sacrificing accuracy when format controls are applied.
Beyond transcription, Sarvam Audio adds diarisation, or speaker labelling, to recordings up to 60 minutes. The model assigns each segment to the correct speaker, a feature aimed at meetings, interviews, and customer service calls where who said what matters as much as the text itself.
The company measures performance using word diarisation error rate and diarisation error rate, claiming state-of-the-art results on in-house meeting data with up to eight speakers and overlapping speech.
A notable addition is direct speech-to-command execution. Instead of transcribing audio first and then passing it to another model, Sarvam Audio can infer intent and extract parameters directly from speech.
For example, a Tamil user asks to pay an electricity bill and the system automatically fills in the biller name, account number, amount, and confirmation, then triggers the payment function. This cuts latency and simplifies production systems for voice bots.
The launch comes as voice interfaces expand across banking, government services, logistics, and creator tools in India. By focusing on code-mixing, script variation, and long conversational audio, Sarvam is positioning its model as infrastructure for “voice-first” apps rather than a basic dictation tool.
Sarvam co-founder Pratyush Kumar recently announced that the company is going on a two-week announcement spree. The company also launched Sarvam Dub, a real-time voice dubbing model, and showcased it during the Union Budget on Republic TV.
Sarvam was the first company to be selected under the IndiaAI Mission to build sovereign foundational models. The company is expected to launch its first model at the India AI Impact Summit 2026 to be held in New Delhi later this month.
Your reaction