cyborg · speech lab

Speech-to-text only, measured. Nothing here reaches the agent, the server, or any model that answers — the microphone, the voice detector and the transcriber all run on this device.
← search

Backend

reading…
Saved for this browser and used by live mode too. Takes effect on the next load, because a pipeline is built once per backend. Force WASM if transcripts come back fluent but wrong — that is what a bad dtype on WebGPU looks like, and it never throws.

Endpointing

How long the detector waits before deciding you have finished. Every millisecond here is dead air the speaker sits through — and every millisecond you take off is a chance of being cut off mid-sentence. There is no correct value, only the one that feels right to you, which is why it is a slider and not a constant. Changes apply to the running mic immediately.

Listen

Idle.

Live flow runs the real loop, minus the agent: it waits for “Hey Cyborg” on the small model, then listens on the full one and shows what it heard, then falls back to waiting after ten seconds of quiet. The wake phrase is the whole utterance — the request is always the next one. It calls the same wakeflow decision live mode calls, so what happens here is what happens there.

Results