Shunya Labs Vāķ real-time translation model launched

Shunya Labs has introduced Vāķ, a real-time translation system covering all 55 Indian mother tongues and enabling 2,970 language pairs. The Shunya Labs Vāķ real-time translation model was unveiled at the India AI Impact Summit 2026 in partnership with Nasscom.

The platform is positioned as an open-weight voice AI release that combines automatic speech recognition, neural text-to-speech and Any-to-Any real-time translation. All model weights are publicly available for download and local deployment.

Addressing India’s linguistic scale

India has more than a thousand language varieties, with 61 languages spoken by over a million people. Many speech AI systems currently support only a limited set of Indian languages, often between five and ten.

According to Shunya Labs, this restricted coverage has led developers, startups and government institutions to depend on foreign APIs and overseas servers. Such dependency places voice data under external terms while leaving large linguistic communities underserved.

Several widely spoken languages, including Bhojpuri, Rajasthani, Chhattisgarhi and Magahi, have seen limited representation in mainstream AI systems. The Shunya Labs Vāķ real-time translation model aims to close this gap by providing comprehensive coverage aligned with India’s linguistic diversity.

Real-time Any-to-Any translation

At the core of the system is a real-time Any-to-Any translation engine. It converts speech from any one of the 55 languages into any of the other 54, resulting in 2,970 possible translation pairs.

Key technical capabilities include:

• End-to-end latency under 1.5 seconds

• Preservation of speaker voice, tone and emotional register

• Zero-shot voice cloning without prior training data

Rather than translating text alone, the system retains vocal identity and expressive nuances across languages.

Three model families released together

The launch includes three distinct but integrated model components.

Pingala ASR, the speech recognition model, is ranked first on the Hugging Face OpenASR Leaderboard. It achieved a Word Error Rate of 3.10 percent and operates with sub-250 millisecond latency on CPU-first architecture, enabling edge and offline deployment.

The neural text-to-speech engine supports all 55 languages with streaming capability, custom voice creation, zero-shot cloning and control over prosody and emotion.

The real-time translation layer integrates these components to deliver low-latency, voice-preserving multilingual interaction.

Sovereign deployment and data control

The Shunya Labs Vāķ real-time translation model is released as open-weight software. Organisations can deploy it on their own infrastructure without external API calls or data leaving their environment.

This approach addresses concerns around citizen voice data moving outside national boundaries. According to the company, the platform eliminates recurring API costs and foreign server dependencies.

The system is positioned for use cases such as government services, healthcare outreach in rural regions, judicial proceedings and multilingual education delivery.

Industry and ecosystem perspective

Sourav Bandyopadhyay, Founder and Chief Scientist, Shunya Labs, stated that India should not rely on foreign APIs to process its own linguistic data. He described Vāķ as an open-weight voice AI system built within the Nasscom ecosystem to support sovereign innovation.

Ankit Bose, Head of AI, Nasscom, said the launch reflects the emergence of high-performance, open-weight models from India’s AI ecosystem. He noted that multilingual coverage and local deployment contribute to inclusive AI development.

Coverage across language families

The system spans 43 Indo-Aryan languages including Hindi, Bengali, Marathi, Gujarati, Urdu, Bhojpuri and Rajasthani; seven Dravidian languages including Telugu, Tamil, Kannada and Malayalam; three Sino-Tibetan languages including Meitei, Bodo and Garo; one Austroasiatic language, Santali; and Indian English.

In aggregate, the coverage reaches over 1.17 billion native speakers.

The launch aligns with national digital initiatives including IndiaAI Mission, Digital India, BHASHINI and Atmanirbhar Bharat.

With its full 2,970 translation pairs available at deployment, the Shunya Labs Vāķ real-time translation model introduces a comprehensive multilingual voice AI framework designed to operate at national scale.

Read More: 

AI Impact Summit 2026 in New Delhi: Getting there, the glitches, and the apology that followed

RiskProfiler Asia-Pacific expansion with Viraat Bindra

Qualcomm $150 million AI startup investment in India expands edge AI focus

Altos Computing and Industry.AI industrial edge AI platform expands manufacturing intelligence