Indian artificial intelligence startup Sarvam AI has launched Saaras V4, a new speech-to-text model designed to support all 22 official Indian languages alongside global English. First reported by MarkTechPost, the model is available starting today through Sarvam‘s Application Programming Interface (API) at a rate of ₹30 per hour. The release targets real-time voice processing and multilingual transcription across India’s diverse linguistic landscape.
Technical Specifications and Capabilities
Architecturally, Saaras V4 pairs a dedicated audio encoder with a 3-billion-parameter hybrid state-space decoder. This setup allows the system to process audio streams efficiently. Sarvam AI claims the model supports real-time streaming with a first-token latency under 150 milliseconds, making it suitable for live conversational applications and real-time captioning. Additionally, Saaras V4 features keyterm prompting, which enables users to supply up to 50 custom terms to guide recognition accuracy for domain-specific vocabulary or proper nouns. A single instance of the model can generate five distinct output modes, though details on the specific output formats were not fully detailed in initial reports.
Missing Details and Performance Questions
While the technical specifications outline low latency and broad language support, several key performance metrics remain unannounced. The initial report did not provide comparative benchmark scores, such as Word Error Rate (WER), against existing open-source speech models like OpenAI’s Whisper or commercial competitors operating in the region. Furthermore, the announcement did not specify the hardware requirements for self-hosting or whether on-premise deployment options will be offered alongside the API access. It also remains unclear how the model handles code-switching—the common practice of mixing Indian languages with English within a single spoken sentence—or how performance varies across lower-resource regional dialects compared to major languages like Hindi or Tamil.
Developers interested in testing Saaras V4 can access the model directly through Sarvam AI’s developer platform starting today. As multilingual voice interfaces become increasingly vital for digital services in India, real-world testing will reveal whether Saaras V4 can maintain high accuracy across its full language roster.
Lead image: https://kaboompics.com/ / Pexels
