Sarvam AI Releases Saaras V4: A Speech-to-Text Model for All 22 Indian Languages and Global English
Sarvam AI released Saaras V4, a speech-to-text model for all 22 Indian languages and global English.
Sarvam AI released Saaras V4, an encoder-decoder speech recognition model available via its API as model "saaras:v4". The decoder is Sarvam-3B, a 3-billion-parameter hybrid state-space language model trained from scratch. Sarvam reports the lowest average word error rate on seven English sets and says Kathbath noisy LLM-WER is under half that of Deepgram Nova-3 and GPT-4o Transcribe; results are vendor-reported and not independently reproduced. Weights are not public. Pricing is ₹30 per hour, or ₹45 with diarization, with streaming time-to-first-token claimed under 150 ms.
- Saaras V4 covers all 22 scheduled Indian languages plus global English.
- Decoder is Sarvam-3B, a 3B hybrid state-space model trained from scratch.
- Vendor reports lowest average English WER; noisy LLM-WER under half of Nova-3 and GPT-4o.
- API-only at ₹30 per hour; weights unpublished and self-hosting still covers v3.
- Five output modes and keyterm prompting of up to 50 terms are included.
From the original article
4 min read at marktechpost.com
Sarvam AI's Saaras V4 is a speech-to-text model covering all 22 Indian languages plus global English. It pairs an audio encoder with a 3B hybrid state-space decoder. It adds keyterm prompting for up to 50 terms, 5 output modes from 1 model, and streaming with first-token latency under 150 ms.
By Asif Razzaq · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.