Technology companies and startups are betting big on voice AI models. Experts say Indian firms are more likely to crack it with affordable pricing and local context

When Dheemanth Reddy began building a voice AI model, his goal was personal before it was commercial. He wanted his parents and grandparents to interact with artificial intelligence in their native tongue, Telugu.

“When AI development took place in the US, what was missing was that 90% of the world still spoke a different language than English. The US and Europe do not represent the entire world, and the only medium for the Global South such as India to access this technology is voice,” said Reddy, cofounder of Maya Research.

Maya Research is among a handful of Indian-founded companies building speech AI models from scratch, alongside startups such as Pixa AI and Soket Labs. Voice AI is increasingly being described as the next frontier in artificial intelligence, with both global technology giants and Indian startups investing heavily in the space.

According to Rajan Anandan, managing director at Peak XV Partners, 2026 could be the year of voice AI for India, with domestic companies gaining ground through affordable pricing and deeper contextual understanding of local languages.

Reddy relocated from New York to India last year to build Maya Research. “Building from the US is like solving India’s problems sitting in the Valley. I must be where the customers are, which is India,” he said.

While acknowledging that voice AI startups are proliferating — “You go into HSR and throw a stone, you will find a bunch of people building Voice AI,” he remarked — Reddy said Maya is focused on building a foundational model supported by its own data collection and labelling teams. The company deploys on-ground data collectors to gather emotional and conversational speech data across dialects and slangs. “The way to teach models conversational and emotional ability is to collect data with multiple speakers and varied dialects,” he explained.

Maya is currently building a 3-billion parameter model, initially in English, with Telugu and Hindi launches planned next. Kannada, Tamil and Malayalam are expected to follow.

At Pixa AI, founder Sparsh Agrawal said the impetus to build arose from dissatisfaction with existing models. “Most voice AI models were bland and lacked emotion. No one model could handle speech and music together,” he said. Pixa has trained multiple smaller language models specialised in music, voice quality and accuracy, combining them into a foundational system.

Other Indian players are also advancing in the space. Sarvam AI has launched a speech-to-text voice model that it claims outperforms existing benchmarks. Meanwhile, global platforms such as ElevenLabs and Cartesia are steadily improving their capabilities in Indian languages including Tamil, Gujarati and Marathi, founders noted.

As demand for voice AI solutions grows, many startups are layering services — such as call centre automation, loan collection and customer support — on top of foundational voice models. Over the past six months, companies report a sharp uptick in demand as model quality improves.

Sudhama Bhatia, cofounder of Travon AI, said no single model currently performs optimally across all use cases. Travon integrates multiple foundational models to power services for hospitals, recruitment firms and collection agencies, including appointment bookings and automated candidate screening.

Bolna AI, founded by Prateek Sachan, similarly orchestrates a mix of Indian and global models, including Sarvam’s, to lower pricing for enterprises. Voice AI services currently cost between Rs 4 and Rs 10 per minute, including telephony and platform charges. Anandan said Indian voice AI firms could potentially reduce costs to around Rs 3 per minute.

However, enterprise-scale adoption remains a challenge. Founders acknowledge that few Indian voice models have yet demonstrated the scale, reliability and robustness of global counterparts such as those from OpenAI or Google DeepMind. Limitations around handling interruptions, application programming interface calls and real-world conversational variability persist.

“The biggest challenge is that even the best global models talk gibberish in Indian languages such as Hindi,” Bhatia said, adding that hallucination remains a persistent issue despite repeated training cycles.

Agrawal highlighted another constraint: talent. “Building voice AI models from the ground up requires exceptional talent, which is expensive and hard to find. The difference between exceptional and great is 100X. Just great is not enough,” he said.

As India positions itself to lead in multilingual voice AI, startups are betting that affordability, contextual depth and on-ground data will give them an edge. Yet scaling to enterprise-grade reliability — and competing with well-funded global labs — remains the defining test ahead.