Type three words — calm elderly school teacher — and a voice reads your text back in exactly that voice: warm, unhurried, a little gravelly. Sprinkle in a <laugh> and it chuckles mid-sentence; write <whisper> and it drops to a whisper. That is Maya1, an open-weight voice AI built by two 23-year-olds in Bengaluru that now ranks above Google on global speech quality leaderboards.
Maya Research was founded in 2025 by BS Dheemanth Reddy, its chief executive, and Bharath Kumar Kakumani, its chief technology officer. Reddy traces the idea to his mother in Vempalli, a small town in Andhra Pradesh: voice was the only interface she could use, yet the speech models of the world did not understand her accent, her rhythm or her expression. Teaching himself English from scratch took Reddy to a master’s at New York University, where he studied deep learning under Yann LeCun — and gave him, as he puts it, a lasting bias toward proving things rather than claiming them.
Maya1 is that proof. A roughly 3-billion-parameter Llama-style transformer with a SNAC neural audio codec, it generates 24 kHz audio with more than twenty inline emotion tags — laugh, sigh, whisper, anger, giggle, sarcasm — embedded directly in the text, so delivery shifts mid-sentence. Instead of fixed voice libraries like ElevenLabs or OpenAI, voices are designed the way you would brief a voice actor: plain-language descriptions such as “calm elderly school teacher.” It streams in real time with sub-100-millisecond latency and runs on a single GPU with 16 GB of VRAM. The weights are released under Apache 2.0 on Hugging Face (maya-research/maya1), free to download, fine-tune and ship.
The reception has been striking. Maya Research says it is the only Indian company on Speech Arena, the blind human-vote benchmark, where Maya1 holds a Quality Elo of 1,051 — sixth among open-weight models worldwide and, per the company, on par with OpenAI’s GPT-Realtime-2. Moneycontrol reports the model sitting above Google on the Artificial Analysis global speech leaderboard. Its successor Maya 2 ranks number one for Hindi on Voice Arena, ahead of the largest labs, and among the top five for English, judged by tens of thousands of blind human votes. The models have crossed 440,000 downloads on Hugging Face, with the company reporting three million app downloads across India, Southeast Asia and the Middle East.
The team did the groundwork others skip: eight months collecting speech across rural India, paying people to record real conversations, tuning the model through thousands of listening sessions. It was built lean — Business Standard reported Maya1 was developed on free cloud credits from AWS and Google Cloud — and the company has raised a $1.9 million seed round from South Park Commons.
The product line is already growing. On mayaresearch.ai, Maya 2 Native is the flagship, built natively in each language to keep a real speaker’s accent and rhythm; Maya 2 Global holds its own on the English board with a 1,089 Elo score against labs many times the company’s size; Maya 2 Flash is the fastest model they make, built for live conversation; and Veena, their first Hindi-and-English model, remains open and downloaded daily. The API streams 24 kHz PCM from regions in Mumbai and Iowa. Next, the company says it is training what it calls a 10- to 15-times-larger Indic speech dataset than currently exists online, targeted for release by June 2026 — with full-duplex, on-device speech-to-speech on the horizon.
More Read
Reddy holds himself to one test: whether his own mother could use what they build, on her own, in her own language. A voice AI that beats Google’s giants from a Bengaluru desk, trained on the voices of rural India — and the interface for the next five billion people is starting to sound like home.
Keep reading: Growing India’s Silicon Backbone: Inside Raana Semiconductors, the Company Building the Machines That Grow Crystals · The Factory That Talks Back: Inside Fynd Forge, Mumbai’s AI-Native Manufacturing OS · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
