Shunya Labs has launched Zero Codeswitch, a speech recognition model aimed at understanding how people in India usually speak, where switching between languages mid-sentence is common.
The Gurugram-based voice AI company said the model is trained to pick up conversations that mix Hindi, English, and regional languages without passing speech through translation layers. According to Shunya Labs, Zero Codeswitch recorded a 3.10 percent Word Error Rate on the OpenASR leaderboard, performing about 48 percent better than the next closest competing system.
The company said the model can run on regular CPU hardware, which helps bring down deployment costs. Shunya Labs estimates organisations could reduce infrastructure spending by as much as 20 times, while still keeping response times under 100 milliseconds. This makes it suitable for real-time use in customer support, voice assistants, and automated call centres.
Built for How India Actually Speaks
Ritu Mehrotra, CEO and co-founder of Shunya Labs, said the company has stayed focused on research rather than chasing short-term attention. She said Zero Codeswitch is part of an effort to build core language technology for Indian use cases, with an emphasis on accuracy, speed, and day-to-day usability. The idea, she added, is to build AI systems locally instead of adapting models developed elsewhere.
Unlike many global speech systems that start with English and are later adjusted for other languages, Shunya Labs said its models are trained directly on large volumes of Indian speech. The training data covers a wide range of accents, dialects, pronunciation styles, and informal language, which helps the system handle Hinglish and other mixed-language speech more reliably.
Sourav Bandyopadhyay, CTO and co-founder of Shunya Labs, said the company’s name reflects its approach to engineering. He said Zero Codeswitch is built to listen to how people actually speak across India, while remaining practical enough for real-world deployment.
The company said Zero Codeswitch is aimed at enterprise and public-sector users where data privacy is a priority. The model can be deployed on local infrastructure or in isolated environments, allowing organisations to customise it while keeping sensitive data in-house. Shunya Labs added that the system meets HIPAA, SOC 2 Type II, and ISO 27001 standards and supports CPU-first deployments to limit reliance on specialised GPU hardware.
Also Read: OpenAI Set to Test Ads on ChatGPT Free and Go Plans in the US
