Sarvam AI Launches Saaras V4, Advancing Speech-to-Text Across twenty two Indian Languages

Sarvam AI has launched Saaras V4, its most capable speech-to-text model to date, delivering state-of-the-art performance across all 22 Indian languages. The model is designed to strengthen AI-powered transcription and voice applications for India’s diverse linguistic landscape.

According to Sarvam AI, 10 of the supported languages currently have no commercial transcription alternative, highlighting Saaras V4’s focus on expanding access to multilingual speech technology.

The model is also built to handle noisy audio environments, recording less than half the error rate of Deepgram Nova-3 and GPT-4o Transcribe on the Kathbath Noisy benchmark. Sarvam AI further reports that Saaras V4 achieved the lowest average word error rate across seven English benchmarks tested by the company.

Saaras V4 introduces keyterm prompting for accurately handling names and specialised terminology, along with five output formats tailored for voice agents, compliance transcription and audio analysis.

The launch marks another step toward scalable, multilingual AI speech technology for India.