[Exclusive] Gnani.ai to Launch 5 New AI Models, Including Speech-to-Text Model

Bengaluru-based voice AI startup Gnani.ai will unveil a series of new models in the coming weeks as it continues to invest in speech AI and frontier model development.

The company is preparing to launch five new AI models over the next month, founder and CEO Ganesh Gopalan told AIM in an exclusive interaction.

While he did not disclose details of the upcoming releases, Gopalan said the models will be launched under a new ‘Prisma’ series, beginning with a speech-to-text model.

“We are launching four to five more models over the next month or so. This is the first in that lineup. It’s also a new nomenclature for us. We call it the Prisma series,” he said.

Gopalan said the company expects to launch the first speech-to-text model within the next few days.

The startup claims the models have been trained on approximately 14 million hours of audio data and outperform both domestic and international competitors on Indian language benchmarks.

“We are finding better accuracy than any local or international competitor,” the CEO claimed.

He added that the company is seeing significantly higher accuracy for Indian languages compared to existing alternatives.

The development would come months after the company unveiled Inya VoiceOS, a 5-billion-parameter voice-to-voice foundation model at the India AI Impact Summit under the IndiaAI Mission. The model supports more than 15 Indian languages.

Inya VoiceOS was trained on more than 14 million hours of multilingual speech pretraining data and over 1.2 million hours of task-specific speech fine-tuning data. The model also incorporates more than 8 trillion text tokens to strengthen its language understanding and reasoning capabilities.

Unlike conventional voice AI systems that convert speech into text before generating a response, Inya VoiceOS consumes and generates speech tokens without converting them into text.

Gnani.ai’s Prisma launch enters a crowded voice AI market dominated by players such as ElevenLabs, Deepgram, and Wispr Flow.

Global players such as ElevenLabs and Wispr Flow have rapidly expanded their language coverage, with the former’s Scribe supporting transcription across 99 languages and Wispr Flow offering voice dictation in more than 100 languages. 

In India, the most prominent competitor is Sarvam AI, which offers the Saaras family of speech recognition models supporting 22 Indian languages, alongside text-to-speech and translation systems.

Another notable player in the space is Smallest.ai, which is building a suite of speech AI models spanning speech-to-text, text-to-speech, and speech-to-speech applications. The startup's Pulse speech recognition model supports 38 languages, while its broader portfolio is focused on low-latency voice interactions and enterprise voice agents. 

Beyond startups, organisations such as IIT Madras research lab AI4Bharat have built open-source speech recognition models and datasets for Indian languages, while Big Tech companies also offer widely used speech services.

Gopalan said the company continues to invest heavily in speech-to-speech technology, an area where he believes very few organisations globally have achieved meaningful progress.

The company is among the 12 startups selected under the IndiaAI Mission to build foundational AI models. Gopalan said the startup was among the earliest recipients of compute resources under the programme.“We were one of the first few companies that had access to the GPUs.”

He added that the company is currently using the government-backed infrastructure to train its next generation of models.

Gnani.ai’s new models will come at a time when India’s foundation model ecosystem is gaining momentum. Earlier this month, Sarvam AI became India’s newest AI unicorn after raising $234 million in the first close of its Series B round led by HCLTech, valuing the company at $1.5 billion.

Separately, Avataar.ai recently unveiled Varya, a distilled video generation model developed under the IndiaAI Mission, joining a growing list of indigenous AI models being built across language, voice, and video applications.