All posts

Research

How streaming state space models keep voice fast

A look at why we build our speech models on streaming sequence architectures, and what that means for live conversation.

Ivo Petrak

7 min read

Live conversation is unforgiving. A pause that would be fine in a podcast feels awkward on a phone call, so our research starts from one question: how fast can a model start speaking without giving up quality?

Streaming by design

State space layers keep a compact memory of everything they have heard so far. That lets a model process audio as it arrives instead of waiting for a full sentence, which keeps the first word of every reply close to instant.

Quality without the wait

We pair those layers with careful training on conversational speech, so voices keep their warmth and rhythm even when they are generated piece by piece.

What is next

We are now exploring hierarchical models that read raw audio directly. Expect more research notes on this blog as that work matures.

Get started today

Talk to an expert.

Meet someone from our team and see how Halvora can power voice experiences your callers will remember.

Start building.

Spin up a voice agent or call our models from the API, and go live the same day.

We’re hiring!

Help build voice that feels human

Create a free website with Framer, the website builder loved by startups, designers and agencies.