Shunya Labs launches Vaķ for real-time translation across 55 Indian languages
Shunya Labs announced the launch of Vaķ at the India AI Impact Summit, an open-weight real-time translation system supporting 55 Indian languages with under 1.5-second latency. The system features any-to-any speech translation, zero-shot voice cloning, and custom voice creation, with an accompanying open-weight speech recognition model and a neural text-to-speech engine. The full model weights are publicly available for on-premises deployment.
Key Takeaways
- Vaķ supports 55 Indian languages, including 43 Indo-Aryan and 7 Dravidian varieties, reaching an estimated 1.17 billion speakers.
- The system delivers end-to-end translation latency under 1.5 seconds and features a CPU-first architecture for edge or offline deployment.
- Shunya Labs' Pingala ASR model recorded a 3.10% word-error rate, the lowest on the Hugging Face OpenASR leaderboard as of the launch.
- Full model weights, neural text-to-speech engines, and speech recognition tools are publicly available for on-premises deployment.
- Zero-shot voice cloning capabilities allow the system to preserve a speaker's emotional tone and voice identity without prior training data.
Why It Matters
This launch establishes a high-performance open-source alternative to proprietary translation APIs, prioritizing data sovereignty through on-premises deployment options. By achieving sub-1.5-second latency on standard hardware, Shunya Labs reduces the technical barriers for localized service delivery in sectors like healthcare and governance. For the streaming ecosystem, this indicates a shift toward cost-efficient, low-latency localization tools that can process hundreds of regional dialects without external cloud reliance. Watch for the adoption rate among Indian government agencies as they look to transition from the state-run Bhashini platform's paid services toward open-weight alternatives.
Additional Context
The release of Vaķ follows a period of rapid professionalization within India's AI startup sector. According to reportage from CIOL in April 2026, Nasscom’s GenAI Foundry program—of which Shunya Labs is a member—has seen its portfolio's projected annual recurring revenue grow from $3.9 million in FY24 to a projected $35 million for FY26. This growth is driven by a pivot from experimental pilots to enterprise-ready deployments, with over 30 participating startups currently raising a combined $120 million to scale infrastructure.
This development also challenges existing domestic infrastructure. India's government-led Bhashini platform, which supports 22 scheduled languages, announced a transition to a paid service model in early 2024 to monetize its API offerings for private entities, per CDO Magazine. While Bhashini has processed over 4 billion inferences via a vendor-agnostic cloud, the introduction of Shunya Labs' open-weight models allows developers to bypass centralized API costs entirely.
Furthermore, Shunya Labs' success on the Hugging Face OpenASR leaderboard highlights a technical trend toward high-efficiency multilingual models. While established players like NVIDIA and Microsoft dominated early 2025 rankings with models optimized for European languages, Shunya Labs' Pingala model—built on a modified Whisper architecture—claims the top spot for accuracy in the Indic segment. This reflects a broader movement toward 'sovereign AI' where local firms develop specialized models to serve linguistic communities often ignored by global LLM providers.
Read full article at ddnews.gov.in
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source