Notes on listening

Short posts about the part of a voice agent nobody benchmarks: the ears. Every number links to the script that produced it.

A browser VAD without an inference runtime

Why we wrote the Silero forward pass by hand in Rust, what it cost, and what it bought: 0.05 ms per frame, 925 KB, 28 ms to a ready model.

The other person in the room

Every detector counts the television as the user. What it took to train one that does not, which ideas failed, and the number we can honestly claim.

What silence costs a voice agent

Audio APIs bill by the minute and TTS by the character. A listener on the device changes what reaches the bill.