listening

A voice arrives as one signal carrying two things at once: who is speaking, and what is being said. They are wound around each other. Nearly every interesting problem in speech is some version of pulling them apart, or of putting them back together on purpose.

I built a low-resource text-to-speech — codec-based, with VAE control over the voice — and got it working on twenty to thirty hours of audio per language and speaker. Most systems expect far more. Working under that constraint is the point.

Alongside it, content-disentangled voice representations — embeddings that hold on to speaker characteristics while leaking as little linguistic information as possible. The leak is the hard part. It is rarely where you expect it.

And full-duplex conversational speech, where both sides can talk at once and neither waits for a turn — which is how people actually speak to each other. That work is currently under review at IEEE SLT.

Indic language identification models that reached state-of-the-art, and ASR trained for production deployment. Enough time inside SLAM, HuBERT, MMS and the generative speech architectures to have opinions about them, including some experimentation well outside the standard training recipes.

How little audio does a voice actually need?

Is full-duplex a modelling problem, or an interface problem wearing a modelling costume?