Shunya Labs Co Founders Ritu Mehrotra and Sourav Bandyopadhyay explain how the model is built to understand natural code mixed speech at scale
Indian speech patterns, where languages blend fluidly within a single sentence, are shaping a new direction for voice AI with the launch of Zero Codeswitch by Shunya Labs. The foundation model has been developed to recognise how people across the country actually speak, without forcing conversations into single language structures or translation pipelines.
Unlike most global speech recognition systems that are optimised for one language at a time, Zero Codeswitch is designed for naturally code mixed speech. In everyday Indian conversations, Hindi, English, and regional languages are often combined using informal grammar, transliteration, and local expressions. These patterns have historically resulted in higher error rates for voice systems, particularly outside major urban centres.
Zero Codeswitch processes mixed language speech natively, removing the need for intermediate translation layers that convert Indian languages into English. Instead, the model uses a unified architecture capable of generating mixed Hindi and English tokens within the same utterance. This approach allows it to accurately transcribe phrases commonly used in daily speech, reflecting how language is used in practice rather than how it is standardised.
The model builds on Shunya Labs’ earlier research milestones, including its performance on the OpenASR leaderboard, where it recorded a word error rate of 3.10 percent. Zero Codeswitch is engineered to run efficiently on standard CPU infrastructure, reducing deployment costs while maintaining sub 100 millisecond latency for real time applications. This design makes it viable for large scale use in environments where access to specialised hardware is limited.
Speaking about the model’s development, Ritu Mehrotra, Co Founder and Chief Executive Officer of Shunya Labs, said the company’s focus has been on foundational research rather than adapting existing global models. She noted that Zero Codeswitch was built to prioritise accuracy, latency, and usability in real world Indian contexts, where voice often serves as the primary digital interface.
Sourav Bandyopadhyay, Co Founder and Chief Technology Officer of Shunya Labs, said the model reflects the company’s first principles approach to building AI systems for Indian languages. He explained that Zero Codeswitch has been trained on millions of hours of Indian audio, capturing variations in accent, dialect, pronunciation, and slang that are often absent from international datasets.
Zero Codeswitch is expected to support voice based applications across sectors such as fintech, healthcare, and public services, particularly in regions with limited English proficiency. The model is designed for enterprise and public sector deployment, with options for on premises and air gapped environments to address data privacy and regulatory requirements.
Headquartered in Gurugram, Shunya Labs is part of a growing ecosystem of companies developing domestic intellectual property in core AI technologies. With Zero Codeswitch now available for enterprise pilots and production use, the company aims to strengthen the infrastructure layer required for voice driven digital access across India’s multilingual population.
At Prittle PrattleNews, featuring you virtuously, we celebrate the commitment and innovation. Led by Editor-in-Chief Smruti Bhalerao, our platform is dedicated to sharing impactful stories that inspire change and create awareness. Follow us on LinkedIn, Instagram, and YouTube for more stories that matter.
