Sarvam AI launches Saaras V4 multilingual speech recognition model

Indian startup Sarvam AI has released Saaras V4, an automatic speech recognition model built on a 3-billion-parameter hybrid state-space language model. It supports 22 Indian languages, code-mixed speech and noisy audio, offers five output formats and streams with under 150 ms time to first token.

TechStory

Open source

Sarvam AI launches Saaras V4 multilingual speech recognition model

Language identification error rate is 5.22% across 22 Indian languages and 2.9% across the top 10
Five output modes: verbatim, normalised, codemix, translit and translate
Streaming recognition with time to first token under 150 ms
Available via API with Python and Node.js SDKs, plus Vercel, LiveKit and Pipecat integrations

#SarvamAI#SaarasV4

Sarvam AI launches Saaras V4 multilingual speech recognition model

Test: Claude Opus 5.5 is 56% cheaper than Opus 5 at equal prices

US Army awards ARC $4.7M to integrate AI weapon sensors into live training

Google Cloud's RRSI uses regularization to stop agent harnesses overfitting benchmarks

Anthropic Blocks Huawei and Amazon Chips From Opus 5.5 Kernel Development