Indian Voice AI Models are Improving. But They Aren’t the Best Yet

During Kerala’s recent elections, John Kuriakose received a campaign call. He thought he was speaking to a political candidate asking for his vote. The voice on the other end sounded familiar. The conversation was convincing enough that it took him some time to realise he wasn’t speaking to the candidate at all, but to a chatbot replicating the politician's voice.

The experience captures both the promise and the unease surrounding India's rapidly evolving voice AI industry.

Voice models are becoming more natural, expressive, and increasingly capable of speaking Indian languages. Bengaluru-based text-to-speech AI startup Murf AI is among many that have recently launched a new voice foundation model, as Indian and global companies are increasingly competing to improve everything from speech naturalness and pronunciation to latency and multilingual capabilities.

For now, much of the commercial opportunity is centred on automated calling agents, customer support, and enterprise voice automation. But companies are increasingly betting that voice could become something bigger, a way for millions of people to interact with AI tutors, job platforms, government services, and personal assistants.

The Race to Build India's Best Voice AI

Murf AI recently launched Falcon 2, a proprietary voice foundation model that the company says ranks ahead of OpenAI Realtime and ElevenLabs Flash and Turbo on Artificial Analysis' Speech Arena for naturalness. Murf says the model can deliver sub-100 millisecond time-to-first-audio and costs $0.01 per generated minute.

The company has designed Falcon 2 to handle challenges central to deploying voice AI in India, including multilingual and code-mixed conversations, diverse accents, and the economics of operating at scale. The model supports more than 150 voices across 35 languages.

“For years, the most important AI models have largely emerged from the US and other established technology hubs,” Sneha Roy, Co-founder and COO of Murf AI, tells AIM. “Falcon 2 represents a different proposition: a voice foundation model developed by our India-based teams and competing directly with established global models on independent benchmarks.”

But determining which model to choose to apply in personal and enterprise use cases is becoming increasingly difficult. There are just too many of them. 

IIT Madras-incubated Bodhan AI, in partnership with AI4Bharat, recently launched four foundational models for India’s education system. Shunya Labs has introduced a multilingual voice platform supporting 216 Indian languages and dialects. Gnani.ai rolled out its Prisma series speech-to-text model trained to handle complex rural accents. Meanwhile, Sarvam AI earlier launched a 3-billion-parameter audio model to process code-mixed speech and low-quality telephony.

That’s where the role of AI labs becomes increasingly critical.

Voice Arena, a research lab evaluating voice AI through blind pairwise human preference judgements, analysed 18,327 approved judgements across 14 systems as of August 22. Native-speaker raters listen to two clips of the same sentence and judge which one sounds better, while also flagging issues such as robotic delivery and mispronunciation.

Its latest Hindi evaluation placed Cartesia's Sonic-3.6 at the top, followed by Google's Gemini 3.1 Flash, and Maya Research's Maya-2-Native.

But Aaditya Pareek of Voice Arena says the more important question is not who occupies the top spot. “The better question, though, is which system leads for which use case, in which language,” Pareek tells AIM. “Naturalness, pronunciation, coverage, latency, and price are separate axes, and no single system dominates all of them.”

The rankings support that argument. Sonic-3.6 led in Voice Arena's Hindi evaluations for customer support, media and entertainment, and general conversation and AI assistants. Gemini 3.1 Flash, however, led in content creation and education.

The leaderboard is also changing rapidly. Since its launch, Voice Arena's Hindi board has expanded from eight systems to 14 and has been re-ranked six times. In the latest round, the gap between the top two systems was just 27 rating points.

Indian Models are Closing the Gap

One of the most interesting findings is that Indian-built systems appear to have a growing advantage when evaluated on Indian speech.

Across 6,951 head-to-head votes in which an Indian-built and a global voice read the same Hindi sentence, native listeners preferred the Indian system 54% of the time, according to Voice Arena. Indian systems were flagged as machine-like on 15% of clips, compared to 23% for global systems. Severe mispronunciation was flagged on 7% of Indian systems' clips compared to 10% for global systems.

However, global models still occupied the top two overall positions.

“The frontier is global, but the middle of the field now belongs to Indian labs,” Pareek observes.

That could be a crucial distinction as the market evolves. Global models may have more resources and larger research teams, but Indian companies could have an advantage in understanding how people in the country actually speak. According to MarketsandMarkets, India’s voice AI market is expected to grow exponentially—from $130 million in 2025 to $914 million by 2030.

Speech is far more complicated than simply generating a sentence. India's everyday conversations are full of code-switching. A single sentence may move between a regional language like Kannada and English. Users may speak with regional accents, insert English names and terms, and mention phone numbers, ticket IDs, addresses or currency amounts.

Voice Arena designed its evaluation corpus around those conditions. Of its 100 Hindi sentences, 93 contain English words, 83 are heavily code-mixed, and 49 contain identifiers or long numbers.

The findings show why that matters. Mispronunciation was flagged on about 25% of clips containing code-mixed sentences, compared to 16.7% for pure Hindi.

Hindi itself remains far from solved. Severe mispronunciations appeared in roughly 15% of Hindi judgements, compared to around 2% for US English.

“The only reliable test is per language, on native ears,” Pareek notes.

Murf's Falcon 2 is also built around some of those real-world challenges. The company says the model's conversational training includes customer calls, support flows, and booking conversations, while its text normalisation capabilities handle numbers, currencies, dates, addresses, pincodes, and vehicle numbers.

Beyond the Calling Agent

But better voice models do not automatically translate into mass adoption.

For now, automated calls remain one of the clearest commercial use cases for voice AI. Enterprises already deal with large volumes of customer service and sales calls, and AI promises to make those interactions cheaper and capable of operating across multiple languages.

Murf says Falcon 2 can support up to 10,000 concurrent calls with stable latency. 

“The challenge was never simply reducing latency or improving speech quality independently,” notes Ankur Edkie, Co-founder and CEO of Murf AI. “It was engineering a system where both improve together without increasing the computational cost of deployment.”

Pareek agrees that economics will be crucial to scaling voice technology in India. “For India, the last point is decisive: unit economics decide whether voice scales at all,” he notes.

But scale and adoption are not the same. A company can operate thousands of simultaneous AI conversations and still struggle to persuade consumers to use voice as their primary way of interacting with technology.

A person may be comfortable asking a chatbot a sensitive question through text but less willing to speak to it aloud in a shared home, office, or public space.

But voice may not always be the most effective way to communicate with an AI either.

Akhil M A, who converses with AI every day for work, says he still prefers typing his queries. Typing gives him more control over how he frames a prompt and gives the LLM greater clarity about what it is expected to do.

“When you speak a prompt, something can be misunderstood,” Akhil shares. “With typing, I can be more precise about what I want the LLM to do.”

The next phase of India’s voice AI race may depend on more than naturalness scores. Voice could make AI more accessible across education, recruitment, government services and personal assistance—but only if users trust it and build the habit of using it.