Uyir Lab Joins NVIDIA Inception

Uyir Lab has been accepted into NVIDIA Inception, NVIDIA's program for startups building with AI. It is a small step for a young lab, and an encouraging one: it gives us access to compute, training and a wider network of investors and partners as we build AI that understands the languages big models miss.

#What NVIDIA Inception is

NVIDIA Inception is a free program for AI startups at every stage. Members get access to discounted GPU cloud capacity, NVIDIA's technical training, member pricing on NVIDIA software and hardware, and introductions to investors through NVIDIA's venture capital network. There is no equity involved, and joining does not make us part of NVIDIA. It is a program that helps startups move faster.

#Why it matters for what we're building

Our work runs on GPUs. Nila, our Tamil model, was trained from scratch on 9.6 billion tokens using PyTorch on NVIDIA GPUs, on a budget most labs would spend in an afternoon. Building it that way taught us how far careful engineering can stretch a small budget. It also showed us the limits. To train larger models, add speech recognition and text-to-speech for real dialects, and expand beyond Tamil, we need more compute than a bootstrapped lab can buy.

Inception helps with exactly that. Over the coming months we plan to use it for three things:

• Bigger and better models. More GPU capacity for the next version of Nila, and for speech models that understand how people actually talk, in Jaffna, Chennai, Madurai, Toronto and London.
• Uyir Bench. Our native-speaker benchmark tests how well AI understands Tamil as it is spoken: dialects, slang and Tamil-English code-mixing. More compute lets us test more models, including open models we run ourselves.
• Learning from the people who build the tools. NVIDIA's toolkits for speech and language models, such as NeMo and Riva, are a natural fit for the voice agents we want to make work in underserved languages.

#What doesn't change

Our approach stays the same. Every recording in our voice studio is given with consent. Native speakers check every benchmark question before it counts. We train only on data we have the right to use. And we go deep before we go wide: one language done properly, with its community, before the next.

Tamil is our first language, not our only one. Our roadmap covers 16 languages and more than 897 million speakers, from Sinhala and Punjabi to Swahili, Yoruba and Quechua. Programs like Inception make that roadmap more realistic.

#How you can help

The most valuable thing we have is not compute. It is the voices of people who speak these languages every day.

• Record your Tamil. Read a sentence or talk freely for a minute in the Uyir voice studio. Your dialect is exactly what AI is missing.
• Test AI yourself. Try the AI challenge and see whether today's AI understands how you really speak.
• Work with us. If you are an AI lab, a company building voice agents, or a researcher working on underserved languages, write to us at hello@uyirlab.com.

#Frequently asked questions

#Is Uyir Lab owned by or partnered with NVIDIA?

No. Uyir Lab is an independent company. NVIDIA Inception is a program that supports startups with resources and introductions; it does not involve ownership or equity.

#Does this change how Uyir Lab uses contributed voice data?

No. Recordings are used only as described in our privacy policy and the consent each contributor gives. Joining Inception does not share any contributor data with NVIDIA.

#Which languages is Uyir Lab working on?

Tamil first, with 15 more on our roadmap. See the full list on our languages page.