Having built speech models for Indic languages, Gnani.AI has developed full-stack capabilities with its own platforms, applications, and proprietary technology. Bharath Shankar, Co-Founder & Chief Product and Engineering Officer, tells Ojasvi Gupta that large global models are limited because they are not optimised for local requirements.
How do you handle linguistic diversity in India?
To address diversity, we have accumulated petabytes of audio data through proprietary and crowd-sourced channels. This data includes a wide mix of accents, genders, and age groups, which helps in building more accurate and inclusive models.
What is your view on AI regulation versus innovation?
There needs to be a balance. Regulation is important to ensure responsible and explainable AI, but it should not restrict innovation. Instead of limiting applications, the focus should be on enabling access to quality datasets, compute infrastructure, and GPUs so that startups can build effectively.
What models have you recently launched?
We have launched a 14-billion parameter speech-to-speech model at the summit. In addition, we have developed a multilingual speech-to-text model trained on 1.5 million hours of audio data, supporting 14 languages with automatic language detection. There is also a text-to-speech model that supports 14 languages and allows cross-lingual capabilities, such as generating multiple languages from a single voice.
What are your expectations from these models?
The expectation is for these models to be widely adopted across sectors such as education, healthcare, and government services. The aim is to enable large-scale, sector-agnostic usage and help enterprises solve real problems rather than just deploying AI for demonstration purposes.
How do you compare your models with those from, say, a Gemini?
Global models often lack strong support for low-resource Indian languages such as Odia or Malayalam, where accuracy can be quite low. Additionally, these models tend to be large and not cost-efficient for Indian use cases, especially when deployed at scale. Purpose-built models are better optimised for both performance and cost.
What is the key limitation of large global models in the Indian context?
The main limitation is that they are not optimised for local requirements. Their size leads to higher inferencing costs and lower efficiency for specific use cases, making them less practical compared to models designed specifically for Indian conditions.
What kind of support have you received from India AI Mission?
We have access to around 1,500 GPUs, including NVIDIA H100 and H200 systems. These are being deployed through a local data centre in India to ensure that models are built domestically and that data remains secure within national boundaries.