What's this about?
Nari Labs wrote up how they got a text-to-speech model responding in under 50 milliseconds, pushing on the speed-and-cost frontier rather than on voice quality.
What exactly is text-to-speech latency?
It is the gap between deciding what to say and the first sound coming out. Under about 200ms, a conversation feels natural; above that, people start talking over each other. Getting under 50ms means the system starts speaking before it has finished planning the sentence — the same trick humans use.
What's the community saying?
HN 171 points, 44 comments
The discussion is technical and largely appreciative, focused on which parts of the pipeline the savings came from. It pairs naturally with the Skyrim companion item, which has the same constraint.