building in voice ai, for low resource languages, here to get feedback and grow with the community

👋 Hi Product Hunt!

I’m Nitesh (Bhudev), co-founder of Vaani Research Labs, building voice AI from Bengaluru, India.

A question has followed me since my IISc days in 2017, when we ran experiments on borrowed supercomputer time: why does AI speech sound so un-human?

Our answer: humanness isn’t a filter you add to speech. Imperfection is what makes a voice feel real, so it has to be built into the model itself. That belief became Vaani.

Before this, I spent years as a Staff Data Scientist at Hypersonix, working on forecasting engines and inventory simulations. Then I moved from predicting numbers to modelling speech.

What we’ve built in year one:
🎙️ A natively multilingual speech foundation model (code-mix included)
⚙️ 148,000 lines of C++ orchestration for ultra-low latency
📦 50 TB of hand-curated low-resource language data
🏦 Production voice agents for banks, lenders and insurers

Off-duty, I play the flute and run after two kids. 🎶

I’d love your take:
1️⃣ What’s the biggest gap in voice AI today: accents, latency, cost, or trust?
2️⃣ Which language or dialect should voice AI get right next?
3️⃣ Developers, researchers, enterprises: what would make you try a new speech stack?

Open to feedback, blogs, products and partnerships. 🙏
vaaniresearch.com