Copyright thehindubusinessline
Pixa AI has launched Luna AI, a speech-to-speech foundational model that doesn’t just talk, but can sing, whisper, pause, and respond with emotional intelligence in real time. Unlike traditional systems that convert speech to text and back, Luna AI is said to directly process audio and generate human-like speech, removing conversion latency and enabling faster, more expressive, emotionally intelligent conversations. Luna can also whisper, pause, sing, and respond in context, creating a level of emotional nuance and responsiveness that feels human. “The current market has two key components: speech-to-text models and text-to-speech models. But Luna is a speech-to-speech, model, meaning it’s speech in and speech out. Luna is built on an emotional first use case because we believe that the future is artificial emotional intelligence. Every time you talk to an AI, it should feel conversational rather than robotic,” explained Sparsh Agrawal, the founder of Pixa AI. Internal evaluations show that Luna outperforms leading real-time systems, including OpenAI’s, on key benchmarks of accuracy and speech naturalness. In Automatic Speech Recognition (ASR), Luna achieved an error rate of 5.24 per cent, surpassing Deepgram Nova’s 8.38 per cent and ElevenLabs Scribe’s 5.81 per cent. In Text-to-Speech Word Error Rate (TTS WER), Luna recorded just 1.3 per cent, outperforming Sesame at 2.9 per cent and GPT-4o TTS at 3.2 per cent. On Mean Opinion Score (MOS) for naturalness, Luna scored 4.62, topping GPT-real-time’s 4.15. Speech-to-speech AI is opening new frontiers in entertainment, mobility, wellness, and companionship. “We are working closely with a few European companies, primarily in the entertainment sector. One is an automobile company that wants to create an AI-powered infotainment or entertainment system for its cars. Another one is a US-based company building AI toys. We see a huge market potential for companies with entertainment-first use cases,” he noted. Other applications include mental health counseling and companionship for older people. Education for kids is another industry the company is evaluating. The founder added that a large company automating customer calls has partnered with them for a proof of concept (POC). Early data from the pilot indicates higher customer engagement, with an increase in call volume and improved conversion rates compared to previous cold-calling efforts. Through a licensing-led business model, Pixa AI aims to make Luna AI’s capabilities available to global partners, enabling applications across entertainment studios, wellness platforms, automotive, and gaming sectors, forming part of its future roadmap for international expansion. “We are starting with B2B applications and, over time, might enable a conversational AI for Indians,” Agrawal noted. Luna has been trained on over millions of hours of speech data, fine-tuned for real-time performance, emotion recognition, and expressiveness. The company has used a mix of synthetic data and open source data to train the model. Currently, Luna can converse in English and understands tonal and dialectal variations across different geographies. Pixa AI plans to introduce multilingual support within the next two to three months, covering 12 major Indian languages and additional global ones, bringing the total to over 30. The startup has raised funding in the single-digit million range and is backed by Nikhil Kamath, Kunal Shah, and Kunal Kapoor, among others. Pixa’s core team currently comprises four full-time members, with plans to expand to about 10 in the coming months, primarily on the technology side. The company also intends to actively participate in the IndiaAI Mission. According to the founder, discussions with government officials have been encouraging, with awareness around GPUs, model development, and data sets. To scale Luna’s multilingual capabilities, Pixa is in talks with the government to secure GPU access under the mission. Published on October 30, 2025
