Call
Stoppers on tool calls ("let me take a look")
"Hmm." if the first token is > 900 ms
Known caller (greets by name)
System prompt for this industry
Don't like her voice?
The pipeline pins the voice to this industry's greeting clip. Re-roll it — a fresh random Fish voice; wait ~15 s for the stoppers to re-warm, then call again.
How to read the numbers
first audio — from the moment you stopped talking (our VAD / xAI's) to the first sample of her voice. Under ~800 ms feels human; 800–1500 needs a stopper; over 2 s is the awkward zone.
stt — batch transcription of your utterance. first token — brain time-to-first-token after STT. tts ttfb — synthesis start.
Barge in any time — she stops. Say goodbye to end; she hangs up.
Turns
no turns yet
| # | first audio | stt | 1st token | tts | tools |
|---|
Session
—