Eminence Voice Agent

Calls 0 live

0

no calls yet

Turns answered —

0

avg reply —

Response time · p50 —

—

caller stops → agent speaks

Success rate 0 errors

—

of turns without an error

Live output NO CALL

start a web call or dial the phone number

-30s-20s-10snow

Response time per turn

STT decision LLM first word TTS first audio target

Pipeline

avg per stage

Live transcript

no call

Errors & events

Network link health of the call you are following

—

Caller audio in · kbps (last 60 s)

Jitter · ms (last 60 s)

measures this browser ↔ this server (through your tunnel), not the caller's phone line

Recent sessions

View all

All sessions

Transcript

select a session

Errors & events

pause

Models

web calls use these now · “Save” makes them the phone default

LLM Speech-to-text Text-to-speech Voice Speaking pace Emotion (Inworld voices) Expressiveness (Sarvam “temperature”)

Turn-taking when the STT decides you have finished

End-of-turn confidence Min silence when confident Max silence Voice activity detection — stop the agent the moment the caller starts talking Interrupt after the caller has been loud for Caller focus — turn down background voices (TV, people behind the caller) before the speech-to-text hears them Ignore voices quieter than the caller by

Phone calls are always 8 kHz, so a higher sample rate cannot separate voices; this uses loudness, because the caller is the one close to the microphone. Smaller number = stricter (a very soft caller may be turned down too).

Lower = cuts in faster but a cough or click may stop the agent. Quiet sounds (like the agent's own echo) are ignored while it speaks.

Lower = the agent answers sooner but may cut you off mid-sentence. Deepgram never ends a turn sooner than 400 ms. Sarvam Saaras is listed but not wired in yet.

Phone line

Twilio voice webhook—