Dheemanth Reddy Is Building the Voice Interface His Family Never Had

A childhood spent watching capable adults struggle with English-first technology shaped his work on multilingual voice AI.

Long before Dheemanth Reddy began building artificial intelligence models, he watched his mother sit beside him while he used a phone. She would tell him about the things she wished she could do on her own, such as ordering something or downloading content. She was educated and curious, but the device expected her to operate in a language and format that did not come naturally to her. The difficulty came from the interface, not from any lack of ability.

Reddy grew up in Vempalli, a small town in Andhra Pradesh, where that experience extended beyond his household. His father and other people around him wanted to use technology but struggled with systems that demanded English and familiarity with unfamiliar menus. Over time, some simply stopped trying. Reddy saw capable people being excluded by tools that had never been designed around the way they communicated.

“The assumption was that they could not use technology,” Reddy says. “That was never the problem. The technology was asking them to leave their own language before they could even begin.”

That became the foundation of his work in voice artificial intelligence. Reddy builds models designed to speak with emotional range, understand what a person means, preserve conversational context, and communicate across languages. He wants users to focus on what they are trying to say rather than on how to operate the software. In his view, speech can remove layers of interaction that have kept many people at a distance from technology.

He did not see a better keyboard or a simpler menu as the answer. He wanted those layers to disappear so users could speak naturally and be understood.

“I never wanted to teach my mother how to navigate a system that was not designed for her,” he says. “I wanted her to be able to speak the way she already speaks and have the technology understand.”

Reaching the point where he could work on that problem required Reddy to cross some of the same barriers he hoped to remove. He grew up in a middle-class family and learned English as part of creating opportunities beyond his hometown. His parents did not have significant financial resources, but they encouraged him to pursue ambitions that extended far beyond what was familiar in Vempalli. Their support helped carry him toward a field that had once seemed distant from the place where he began.

That path eventually took him to New York University, where he earned a master’s degree in Computing, Entrepreneurship, and Innovation. He studied deep learning under Yann LeCun, whose work helped shape the modern field of artificial intelligence. The experience gave Reddy the technical grounding to pursue a question that had followed him from home. He began to see that the problem he understood personally could also be approached through model development.

“At home, I understood the human problem,” he says. “At NYU, I began to understand how difficult the technical problem was and why it was worth solving.”

Speech generation had already advanced considerably, particularly in English, but Reddy saw weaknesses beneath polished demonstrations. Many systems could read written text aloud while still struggling with emotion or native-language quality. Conversational systems also had difficulty retaining meaning across an exchange. Those shortcomings became more visible outside English, where high-quality data was often limited or unavailable.

Reddy chose to make human-sounding speech the center of his work. That required him to address pronunciation, rhythm, emotion, and conversational understanding as connected parts of the same experience. A model could not sound natural simply because it pronounced individual words correctly. It also had to reflect how people actually speak and respond within a conversation.

“People know immediately when a voice does not belong to their language,” he says. “The words may be technically correct, but something in the rhythm or emotion tells them the system was built somewhere else and adapted later.”

The work led Reddy to build Maya 1, an open-weights speech model that shipped with 21 distinct emotions. It could produce laughter, whispers, anger, and other forms of expression that help spoken language carry meaning beyond the words alone. Maya 1 has received more than 330,000 downloads and has been used by developers and companies worldwide. Its reach gave Reddy a public measure of whether the model was useful beyond a controlled demonstration.

Maya 1 has also established Reddy’s work beyond its download count. The model is one of the world’s top open-weights speech models and the only model from India represented on the Artificial Analysis / Speech Arena leaderboard. Reddy has also been invited to evaluate other technical work as a judge at the Cerebral Valley AI hackathon, extending his involvement in the field beyond the models he builds himself.

Developing multilingual speech also forced him to confront the shortage of usable data for many languages and dialects. Existing resources were not always strong enough to support the level of native quality he wanted. Reddy collected real speech and worked dialect by dialect to create a better foundation. The process reinforced his belief that multilingual systems cannot be built well by treating language as a secondary addition.

“A language cannot be added as a label at the end,” he says. “You have to understand how people actually speak, where they pause, and what feels natural to someone who has heard that language their entire life.”

Reddy later built Maya 2, which reached the number one position for Hindi on Voice Arena. The ranking was based on more than 15,000 blind human votes and placed the model ahead of ElevenLabs and Cartesia on the Hindi leaderboard. For Reddy, the result mattered because listeners judged the voices without being asked to accept claims from the people who built them. It offered outside evidence that a model designed around native-language quality could perform at a high level.

He has consistently favored public testing over controlled demonstrations. Maya 1 was released as an open-weights model, while Maya 2 was tested through blind listener comparisons. That approach reflects a broader professional philosophy built around independent proof. Reddy believes strong work should be able to withstand evaluation beyond the team responsible for creating it.

“Proof has to come from outside the room where the model was built,” he says. “You can create a perfect demonstration, but real people decide whether a voice sounds natural to them.”

Reddy now serves as co-founder and CEO of Maya Research, which is backed by South Park Commons. He also received an Emergent Ventures grant for his work on voice AI that sounds native. These milestones support a mission that remains closely tied to the people who first revealed the problem to him. His work continues to focus on making spoken interaction feel natural across languages that have often received less attention from the technology industry.

He believes voice will become a primary way people use technology because speech removes barriers that screens and keyboards preserve. A person should be able to ask for help or complete a task by speaking in the language used at home. The system should understand the words, emotional cues, and context without requiring the user to translate first. That future, in Reddy’s view, begins with designing for the people technology has overlooked.

His measure of progress remains personal.“My test is still whether my mother could use what I build,” he says. “When she can speak naturally and be understood, the technology is finally doing what it should.”