The Voice Lab
The Voice Lab sits at Phone → Setup and is open to both audiences. Every control on it declares which of two kinds it is: an engine setting that becomes a session parameter and takes effect exactly, or a steering setting that becomes a line in the instructions and is interpreted. A draft can be heard on a test call before it is saved.
The honesty the screen is built on#
The realtime engine is speech-to-speech. There is no text-to-speech stage underneath it, so there is no synthesiser with a pitch dial to reach — and a tuning screen that pretends otherwise would be offering controls that quietly do nothing. Rather than hide that, every control names its kind.
| Kind | Where it goes | What it does |
|---|---|---|
| Engine | voice_style.session_options, as a session parameter | Exactly what you set |
| Steering | The instructions, via voice_style.style_block | A direction the model interprets |
It is a small labelling decision with a large effect on how people use the screen. Somebody adjusting a steering control knows to judge it by listening to a call; somebody adjusting an engine control knows the value is the behaviour.
What is on each side#
- Engine settings
- Model, voice, a pinned language code, temperature, reply length, affective dialog (native-audio models only), the model's end-of-speech silence (
fast= 450 ms), pre-emptive generation, thinking off, the interruption thresholds in seconds and words, false-interruption resume, noise cancellation, and the silence-check and hang-up timers. - Steering settings
- Speed, pitch, energy, warmth, emotion, pauses, fillers, formality, humanness, accent and pronunciation.
- Behaviour texts
- Role, goal, tone, style, knowledge, rules, what to ask and never ask, objections, escalation, opening, closing, proactiveness, language switching and memory use — rendered by
behaviour_block.
Pitch appearing on the steering side is the clearest illustration of the labelling. You can ask for a higher, lighter delivery and often get it; you cannot set a frequency, and the screen does not suggest you can. The related parameters that genuinely cannot be expressed at all — Hz, dB, jitter, formants, exact pause lengths, a delay before answering — are shown as unsupported in human voice profiles.
Drafts, test calls and the two refusals#
A draft is not saved configuration. It sits at the end of the resolution order — after workspace, default, purpose, line and contact — and applies to a test call so a change can be heard before any real caller meets it.
Two refusals happen before a test call becomes a row, and both exist because of a specific failure:
bad_model— the model id is a Vertex path or otherwise malformed. The fifth Lab call rang the owner's phone and then died with a 1008; refusing before dialling is better than a phone that rings into a broken session.not_ready— no worker has checked in for 90 seconds. Refusing here is what lets the drain retry rather than leaving a call attached to nothing.
After the call, the resolved settings are snapshotted onto call.quality. That snapshot is what the call quality review judges, so a call is never scored against settings that were changed afterwards.
Working order#
Change one thing. Prefer an engine setting when one exists for what you want.
Result An engine change is attributable; two steering changes at once are not.
Place a test call with the draft.
Result You hear the change, and the settings that produced it are recorded on that call.
Read the review before saving.
Result If the finding you were chasing is marked model-limited, no amount of further tuning will move it — and the review says so instead of implying a control.
Questions#
I changed a steering setting and heard no difference. Is it broken?
Probably not, and that is the point of the labelling. A steering setting is a sentence in the instructions; the model weighs it against everything else it has been told. If a behaviour matters enough to be guaranteed, look for an engine setting or deterministic text instead.
Can I tune settings while a call is happening?
Not from the Lab — that is what the supervisor panel is for, and it steers behaviour with guidance rather than settings. On the default model even that cannot be delivered, because the SDK marks its session immutable.
Does a test call cost the same as a real one?
It is a real call on a real session, so yes — it is metered like any other, on audio tokens, and it is subject to the same budget refusal. There is no free rehearsal mode, which is a good reason to change one thing at a time.