The race to build home-grown foundational artificial intelligence (AI) models is intensifying, with several startups building sovereign models. Sarvam AI said on Wednesday that its homegrown foundational large language model (LLM) could debut in early 2026. The announcement comes after Jaipur-based Pixa AI, on October 29, unveiled Luna, a speech-to-speech foundational AI model that laughs, sings and even sneezes.
Founder Sparsh Agrawal (25), began work on the project only in March 2025, trained it on borrowed GPUs and cloud credits, and is already in talks with companies to commercialise it. “We didn’t have a huge research lab or a $100 million runway,” he says. “Luna is proof that world-class technology can be built from India with resourcefulness.” The project is backed by Nikhil Kamath’s WTFund, Cred’s Kunal Shah, PhysicsWallah’s Prateek Maheshwari and actor Kunal Kapoor.
What sets Luna apart is its architecture. Unlike most systems that convert speech to text and back, Luna processes sound directly— speech-to-speech—producing a new voice output that carries tone, mood and rhythm. Luna can laugh, talk in accents, pause mid-sentence and even sing. “We wanted to build something that understands how you speak, not just what you say,” Agrawal explains. “Emotion is the missing piece in human-machine interaction.”
Kalika Bali, senior principal researcher at Microsoft Research India (MSR), and her team have been working to make AI more culturally adaptive, ensuring language models account for regional nuances across the globe. “Bias does not translate," says Bali, who advocates for India to develop its own LLMs and is working to make AI more inclusive, ethical and human. Pixa AI has the same goal, says Agrawal. And though it is not a part of the IndiaAI Mission, he says the company is in touch with the government.
Also Read: Quantum Dreams and AI realities: Sundar Pichai’s next big leap
Alongside LLM developers such as Sarvam AI and Krutrim, Pixa AI joins the likes of Gnani AI among voice-focussed players. Gnani, backed by the IndiaAI Mission, is training multilingual voice systems for real-time enterprise use. Luna, meanwhile, positions itself as emotion-first, with a model designed not just to understand meaning, but to perform feeling.
Agrawal sees that distinction as strategic. “Most voice AI today is built for automating customer-support centres,” he says. “We want to be the default for entertainment and emotion-driven applications. We are already working with an automobile company to create an AI-powered entertainment system from scratch.”
Since the launch, Agrawal says, the startup has been flooded with inbound interest from platforms in both India and the US. One of the earliest pilots involves using Luna to simulate patient voices for medical-school training—angry, coughing or tearful—so that students can practise bedside empathy safely. “A global company wanted to mimic patients who are very angry or sad,” he recalls. “It’s a funny use-case, but it shows how emotional range can have real-world utility.”
This gap led Kapoor to invest. “What really got me interested in Luna was the fact that Sparsh was not only trying to build technology, but he was also trying to solve a storytelling problem,” he says. “A lot of the AI we have today is very robotic. You don’t feel like you’re talking to a human being. So how do you create something that has a greater emotional quotient, that understands cultural context and the nuances of language?”
Kapoor says he is deeply involved with the project. “I don’t want to be a person that just writes the cheque and disappears. So, with Luna, I’m an investor, a strategic advisor, a mentor when I need to be, and I'm also a cheerleader.” For him, the opportunity is emotional as much as commercial. “If AI is going to talk to people, especially children, it must also connect with them,” he says. “That’s what makes Luna special. It’s not just answering questions; it’s teaching, comforting, storytelling.”
First Published: Nov 20, 2025, 15:48
Subscribe Now