Gurugram, May 11: Shunya Labs’ open-weight, real-time voice AI system, Vak, supports 55 Indian languages and is built for enterprise and public-sector deployment. Trained on all annotated portions of Project Vaani, Vak combines large-scale Indian speech data with production-grade voice AI infrastructure into a deployable multilingual system. It addresses a longstanding gap in India’s voice AI ecosystem, where developers and enterprises have often had to compromise between multilingual capability, deployment flexibility, and data sovereignty.
Instituted in 2022 by the Indian Institute of Science, ARTPARK, and Google, Project Vaani was created to capture how India actually speaks at scale. The dataset today includes 31,255 hours of spontaneous speech from 156,534 speakers across 165 districts, 31 states and union territories, and 109 languages. The dataset also comprises 288,429 Images & 2043 Hrs of text, making it the largest open-source multimodal corpus for Indian languages ever built.
Sourav Bandyopadhyay, Founder and Chief Scientist at Shunya Labssaid, “India does not need to depend on foreign APIs to hear its own people. With Vak, we are releasing a real-time open-weight voice AI system supporting 55 Indian languages and built for how India actually speaks. Developers can build on it, enterprises can deploy it within their own infrastructure, and users can interact in the languages they use every day.”
Vak’s real-time speech translation layer supports any-to-any translation across all 55 supported languages with end-to-end latency under 1.5 seconds while preserving speaker voice and emotional tone.
The system was trained on spontaneous speech collected under real-world acoustic conditions, including regional accents, background noise, and code-switching between languages such as Hindi and English within the same conversation. The training approach reflects India’s linguistic reality rather than studio-recorded benchmark conditions.
Dr. Prasanta Kumar Ghosh, Professor at IISc said, “Project Vaani was created to make diverse speech data available so that language technologies can be built for a wider set of users. Shunya Labs’ work on Vāķ shows how such data can be used to build systems that work across languages.”
Complete model weights are publicly available for local deployment, allowing organisations to run Vak entirely within their own infrastructure without routing voice data through external servers. The system is designed for government agencies, healthcare networks, financial institutions, and enterprises operating in regulated environments where data control and low-latency deployment are critical requirements.
“Building high-performance open-weight models across Indian languages reflects the growing strength of India’s AI ecosystem. Equally important is the ability to deploy these systems securely within local infrastructure at scale,” Ankit Bose, Head of AI at Nasscom
The release of Vak reflects a broader shift toward sovereign AI infrastructure designed around multilingual accessibility, deployment control, and production readiness rather than benchmark performance alone.
