Skip to main content
Shunya Labs' Zero Codeswitch AI model for Indian code-mixed speech, improving voice AI for 1.4 billion Indians. [indiatechnol

AI news illustration: Shunya Labs launches Zero Codeswitch AI model for Indian code-mixed speech

Zero Codeswitch AI by Shunya Labs Boosts Code‑Mixed Speech

Updated: 3 min read

Most speech AI hears the world in an American accent. Shunya Labs built one that hears India. The company just launched Zero Codeswitch, a model trained on millions of hours of actual Indian speech, not cleaned-up English transcripts.

It understands the fluid, messy switch between Hindi and English, or Tamil and English, that defines daily conversation. The model doesn't translate. It recognizes code-mixing as a primary language.

"With Zero Codeswitch, we are building foundational technology for Indian languages that prioritises accuracy, latency and real-world usability. Our goal is not just to adopt AI, but to build it at the foundation level in India." Unlike global speech models that are primarily trained on English data and later adapted for Indian languages, Shunya Labs said its foundation models are trained from the ground up on millions of hours of real-world Indian speech data. This includes variations in accent, dialect, pronunciation and slang across regions, allowing the system to better handle Hinglish and other code-mixed speech patterns.

Building from the ground up on real data means the model’s advantages are fundamental, not cosmetic. Lower latency happens because the system isn't constantly trying to map sounds back to a foreign phonetic dictionary. Better accuracy comes from training on the target, not an approximation.

This is a practical choice. It prioritizes how people actually use technology over how engineers wish they would.

Shunya Labs is betting that India's digital infrastructure, from voice assistants to call centers, needs a native foundation. The alternative is a future where machines still ask you to speak more clearly, in a language you don't fully use.

Common Questions Answered

What specific challenge in Indian speech does Shunya Labs aim to address with the Zero Codeswitch AI model?

Shunya Labs targets the difficulty global speech‑recognition systems face with code‑mixed conversations that shift between Hindi, English, Tamil and other regional languages within a single sentence. The company argues that this challenge creates a barrier for everyday applications such as voice assistants, transcription services, and customer‑support bots for millions of Indian users.

How is the training approach of Zero Codeswitch different from that of typical global speech models?

Unlike most global models that are first trained on monolingual English corpora and later adapted for Indian languages, Zero Codeswitch is built from the ground up using millions of hours of real‑world Indian speech data. This foundation‑model strategy aims to improve accuracy, latency and real‑world usability for Indian language applications.

Which languages does the Zero Codeswitch model claim to handle within a single utterance?

The article states that Zero Codeswitch can process Hindi, English and various regional tongues—including Tamil and other local languages—within the same spoken sentence. By handling multiple languages simultaneously, it eliminates the translation layers that often hinder existing voice assistants.

What details about Zero Codeswitch's performance metrics have been released so far?

Performance metrics for Zero Codeswitch have not been disclosed; the article notes that accuracy and latency figures remain largely unverified. Consequently, while the model’s capabilities are described in broad terms, concrete quantitative results are still pending.

LIVE03:21OpenAI's Miles Wang in Talks for USD 2B AI Drug Discovery Startup