Developers building voice applications for India often face inaccurate transcription for regional languages, limited support for code‑mixed speech, and the hassle of post‑processing to get the desired output format. They also need low‑latency streaming for real‑time agents, affordable batch processing for long recordings, and a way to bias the model toward domain‑specific terms such as product names or internal jargon.
Sarvam AI’s newest model Saaras V4 solves these problems. It is an encoder‑decoder system that covers all 22 scheduled Indian languages plus English, including global English accents, and reports state‑of‑the‑art word error rates across those languages. The model is available today through Sarvam’s API by simply setting model=”saaras:v4″; no self‑hosting weights are required because the current SageMaker self‑hosting guide only covers V3.
Saaras V4 returns five useful representations from a single audio stream: transcribe (normalized script), verbatim (fillers and spoken numbers kept), codemix (English words stay in Latin script), translit (full utterance in Latin script), and translate (English translation with normalized numbers). This eliminates extra post‑processing steps that can introduce errors.
For domain‑specific vocabulary, V4 introduces keyterm prompting. Users can pass up to 50 terms of 64 characters each; the model biases recognition toward those terms without guaranteeing them, and works best with the codemix mode.
Performance highlights include streaming WebSocket delivery with time‑to‑first‑token under 150 ms, REST for clips up to 30 seconds, and asynchronous batch jobs for files up to two hours with optional speaker diarization. Pricing is ₹30 per hour for real‑time, streaming or batch transcription, rising to ₹45 per hour when diarization is enabled. SDKs are provided for Python 3.9+, Node.js 18+, plus ready‑made integrations with LiveKit Agents, Pipecat and the Vercel AI SDK.
Switching from the previous V3 model is a one‑line change in the request, making adoption fast and low‑risk. Teams can now obtain accurate, multilingual transcripts in the format they need, with low latency and predictable cost, without managing their own infrastructure.
#AI #SpeechToText #IndicLanguages #Productivity #VoiceAI #Tech