Saaras V4: Sarvam AI’s new speech recognition model adds 22 Indian languages, 5 speech formats

Sarvam AI has launched Saaras V4, a new speech recognition model supporting 22 Indian languages and English. The model offers five output modes: transcription, translation, transliteration, verbatim and code-mixed text.

Sarvam AI has launched Saaras V4, a new automatic speech recognition (ASR) model designed for multilingual voice applications. In addition to English, it supports 22 Indian languages and is built to handle code-mixed speech, regional accents, dialect variations and noisy audio. The model is part of Sarvam AI's initiative to expand the use of voice AI in India, particularly across Indian languages. The Bengaluru-based company announced Saaras V4 in a product update titled “Introducing Saaras V4” on 24 August.

Sarvam AI said Saaras V4 supports five types of speech processing: transcription, translation, transliteration, verbatim and code-mixed text. It also supports low-latency streaming, with a reported time-to-first-token of less than 150 milliseconds, making it suitable for real-time voice applications.

Saaras V4 is built for Indian languages

According to Sarvam AI, Saaras V4 combines an audio encoder with a 3-billion-parameter hybrid state-space language model trained in-house from scratch. The company said the model is for real-world audio conditions, including background noise, code-switching and dialectal variations.

Based on verified IndicVoices data, the model recorded a language identification error rate was 5.22% across all 22 languages and 2.9% across the 10 most widely spoken Indian languages.

Five output modes in one model

Saaras V4 can produce multiple versions of the same spoken material without requiring separate post-processing systems.

The modes are verbatim, transcribe, codemix, translit and translate. Verbatim preserves the exact words of the original script, and the transcribe mode will standardize numbers and dates. Codemix keeps English words in English, translit converts speech into Roman script, and then translates it into English.

Pragya Singha Roy