AI and voice models in detail
The panel behind how your agent thinks and sounds: provider, model, response style, transcription and voice. What each control does and which one to reach for.
7 min readUpdated 26 August 2026
The Voice & model card higher up the page covers the three choices most people need. This panel is the full set, and it opens with the most useful thing on it.
The pipeline strip
Across the top sits the path a call actually takes, drawn as three stages with the provider named under each: Speech to Text, then Thinking, then Text to Speech.
It is worth a look before you change anything, because it tells you at a glance whether you are running a standard pipeline with three separate components or a realtime model that does the whole thing itself. That single fact decides which of the controls below will do anything.
AI Model
Described as the AI that powers your agent. Two fields and a slider.
| Control | The page describes it as |
|---|---|
| AI provider | The company whose AI runs your agent |
| AI model | The specific model your agent thinks with |
| Response style | Lower = more consistent answers, higher = more varied |
What Response style actually changes
It is a slider, and it is not a quality dial. It controls how much the agent varies its wording between calls.
Low means the same question tends to produce the same answer. That is what you want for anything where the answer is a fact: hours, prices, policy. Predictability is the feature.
Higher means more variety in phrasing. That reads as more natural in open conversation and it is exactly wrong for a compliance script, where two callers getting differently worded versions of the same disclosure is a problem.
Tip
If your agent occasionally phrases something in a way that worries you, lower this before rewriting the prompt. Variety is what let it wander in the first place.
Speech and Voice
Described as how your agent understands callers and how it sounds back. Four fields and a language.
| Control | What it does |
|---|---|
| Transcription provider | How the agent understands what callers say |
| Transcription model | The specific model doing that |
| Voice provider | Who generates the agent's speech |
| Voice | The voice itself, shown with its provider and gender |
| TTS model | The model producing the audio |
| Language | What the agent expects to hear and speak |
Picking a voice sets three things at once
The page says this directly and it saves a lot of confusion: one pick sets the voice provider, model and voice together, and only priced voices your AI model supports are shown.
Two consequences follow. You do not assemble a voice from parts, you choose one and the rest follows. And a voice you expected to see and cannot is not missing; it is either not priced on your account or not supported by the model you have chosen.
Careful
If a voice disappears from the list after you change the AI model, that is the support filter, not a bug. Change the model back or pick from what is offered.
Transcription is where mishearing lives
The single most common complaint about voice agents is that they misheard someone, and the reflex is to blame the model that answered. The model never heard the audio. It only saw text.
So when a call goes wrong in that particular way, the order is: read the transcript, see whether the text matches what was said, and if it does not, change the transcription provider or the language rather than anything else.
Common questions
- What does Response style do?
- It controls how much the agent varies its wording. Lower is more consistent, higher is more varied. It is not a quality setting.
- Why can I not see the voice I want?
- The list only shows priced voices that your chosen AI model supports. Changing the model changes the list.
- Do I pick the voice provider or the voice?
- The voice. Choosing it sets the provider and the model with it.
- My agent mishears callers. Which setting?
- Transcription provider and language, not the AI model. Check a transcript first: if the text is wrong, the problem happened before the model was involved.
- What is the strip at the top of the panel?
- The pipeline your agent runs, stage by stage, with the provider for each. It tells you whether you are on a standard three-stage pipeline or a realtime model.
Was this helpful?