Gnani.ai unveils Inya VoiceOS, India's first 5-billion-parameter voice-to-voice AI model

India AI Impact Summit 2026: Gnani.ai Launches India's First Voice-to-Voice AI System '5B Inya VoiceOS'

The launch highlights the strong progress made under the IndiaAI Mission and the country's growing focus on artificial intelligence. Inya VoiceOS is India's first 5-billion-parameter model. Most voice systems change speech into text first and then convert text back into speech. Inya VoiceOS works differently by understanding and responding directly to speech. The conversations sound more natural while ensuring a clear tone and emotion. Gnani.ai trained the model using more than 14 million hours of multilingual speech data. The team fine-tuned it with over 1.2 million hours of task-based audio. Training also included trillions of text tokens. The company built and deployed the system in India.

India AI Impact Summit 2026: Gnani.ai unveils India's first 5B voice-to-voice AI model, Inya VoiceOS

Currently in research preview, a larger 14B-parameter version is already in development. India's AI startup Gnani.ai introduced what it describes as the country's first 5-billion-parameter voice-to-voice foundational model under the India AI Mission. The model, called Inya VoiceOS, was showcased at the India AI Impact Summit 2026 at Bharat Mandapam, where Prime Minister Narendra Modi formally unveiled it. Unlike conventional voice assistants that convert speech into text and then back into audio, Inya VoiceOS works directly with speech. The system processes and generates spoken responses natively, without relying on intermediate speech-to-text or text-to-speech layers. According to the company, this allows it to better retain tone, pauses, emotion and other subtle cues that shape human conversations. Gnani.ai said the model has been trained on one of the largest sovereign voice datasets assembled for Indian languages. It is built on 5 billion parameters, pre-trained on over 14 million hours of multilingual speech data and further fine-tuned on more than 1.2 million hours of task-specific audio. The training also incorporated trillions of text tokens to strengthen reasoning and linguistic grounding. The company claims sub-second response times and 24 kHz audio output with natural-sounding prosody. It supports more than 15 Indian languages and is designed to handle code-mixed conversations, a common pattern in India, along with interruptions and overlapping speech during real-time interactions. It can be used for multilingual government helplines, grievance redressal systems and emergency response platforms. For the private sector, it could power hands-free, voice-driven workflows in banking, insurance, healthcare and logistics. It must be noted that the current release is a research preview, with Gnani.ai indicating that a larger 14-billion-parameter version is in development. The company emphasised that the model has been built and deployed entirely within India.