Shunya Labs Unveils Real-Time AI Translation Model for 55 Indian Languages

Gurugram-based startup Shunya Labs has launched Vāķ, an open-weight, real-time voice translation model that supports 55 Indian languages across 2,970 translation pairs. The model was unveiled in partnership with Nasscom at the India AI Impact Summit 2026.

Positioned as India’s largest open-weight voice AI release, Vāķ enables ‘any-to-any’ speech translation across Indian languages, with end-to-end latency under 1.5 seconds. The system not only translates speech but also preserves the speaker’s original voice, tone, and emotional nuance, the company noted. It also features zero-shot voice cloning, allowing voice replication without prior training samples.

Vāķ aims to address a longstanding gap in India’s speech AI ecosystem, where most systems support only a limited set of languages, often pushing developers and institutions to rely on foreign APIs and overseas servers. According to Shunya Labs, Vāķ enables fully sovereign deployment, allowing organisations to run the model on their own infrastructure without data leaving the country.

The release includes three model families: Pingala ASR, with a reported 3.10% Word Error Rate on the Hugging Face OpenASR leaderboard; a neural text-to-speech engine offering emotion and prosody control; and the real-time Any-to-Any translation engine covering all 2,970 language pairs.

Sourav Bandyopadhyay, Founder and Chief Scientist of Shunya Labs, said the company’s goal is to reduce India’s reliance on foreign voice AI systems. “India does not need to depend on foreign APIs to hear its own people. With Vāķ, we are releasing a high-performing open-weight Voice AI across 55 distinct Indian tongues. Every developer can build, every government can deploy, and every Indian can be heard in their own language,” he said in a statement.

Ankit Bose, Head of AI at Nasscom, said the launch reflects the growing maturity of India’s AI startup ecosystem and its focus on sovereign, inclusive AI systems.

Vāķ covers 43 Indo-Aryan languages including Hindi, Bengali, and Marathi; seven Dravidian languages including Tamil, Telug,u and Kannada; three Sino-Tibetan languages including Meitei and Bodo; Santali from the Austroasiatic family; and Indian English. The company claims the system collectively serves over 1.17 billion native speakers.

Shunya Labs said the complete model weights are publicly available for download, enabling enterprises, governments and developers to deploy the system without recurring API costs or external dependencies.