loading
loading
What Indian users consider an ordinary conversation could be the toughest test for voice AI, and solving it could unlock applications worldwide. An AI voice can answer a question, confirm an order or help book a reservation. But can it pronounce an Indian name correctly, understand an address spoken in a mix of Hindi and English, and respond quickly enough to sound natural? For Sneha Roy, the co-founder and COO of Murf AI, these are not exceptional scenarios. They are the baseline for building voice AI that works for India and, ultimately, the world. During a masterclass on ‘Building voice AI from India, for India, and for the world’, at TechSparks 2026, Roy explained how India’s linguistic complexity, infrastructure constraints, and cost considerations are influencing the development of speech technology. When edge cases become the specification Roy drew parallels with products that adapted to Indian users before finding wider applications. Google Maps introduced offline functionality for areas with limited internet connectivity, while Netflix introduced a mobile-only plan that, Roy said, has since been adopted in 78 countries. Uber piloted cash rides in Hyderabad, before expanding the option to other markets. She believes voice AI could follow a similar path. Consider a routine delivery-status call in Bengaluru. A customer might combine English with a regional language, use a locally familiar pronunciation of an address, and expect a PIN code to be spoken digit by digit. A model built around standardised English can struggle with these details. “If the customer has to change something about the way they speak just to get the agent to cope up, or keep repeating their address or their name, we haven’t finished building it yet,” said Roy. For her, the challenge is to build systems that accommodate how people naturally speak, rather than expecting customers to adapt to the technology. Smaller models, faster conversations Natural speech is only one part of the problem. Voice agents must also respond quickly and operate economically at scale, where businesses may handle millions of calls each month. Roy said the complete voice-agent pipeline, including speech recognition, the language model and speech generation, typically has a response budget of 500–800 milliseconds. Murf has worked to bring its text-to-speech model latency below 100 milliseconds, including network delays, she said. To achieve this, the company developed Falcon 2, a smaller speech model designed to run across a wider range of GPUs rather than depend on specific, high-end hardware. Roy said this approach has helped maintain low latency across different geographies while improving cost efficiency. The same engineering choices can benefit markets beyond India, particularly where computing infrastructure is limited or deploying voice agents at scale remains expensive. Building trust alongside natural speech Roy also highlighted the importance of responsible voice data practices. Murf works with more than 400 voice actors globally through long-term, royalty-based partnerships, she said, and does not clone voices without consent. For customers with strict privacy requirements, the company offers APIs that can process data without retaining it, as well as on-premises deployments for sensitive applications. Roy’s central argument is that India’s linguistic diversity and infrastructure constraints should not be treated as secondary requirements. They can help define what voice AI needs to do before it is ready for widespread adoption. As she put it, “India is the spec capital of the world.”
document fetched 2026-10-27 02:57