Voices

Audition candidate voices against your own script, then save the one you want and point a workflow at it.

Choosing a voice is three decisions, not one: a provider, a model within that provider, and a voice from that model's catalogue. You make all three in one place.

• Voice Bench — the audition surface. Type a script, hear it read, tune the provider's settings, and compare two candidates side by side.
• Voice Profile — a saved, named voice, owned by your organization. Profiles are reusable: one profile can serve every workflow that should sound the same.
• Fallback voice — a second configuration stored on the profile, used if the primary one fails to start when a call begins.

Providers

The text-to-speech providers wired into the platform today are Deepgram, OpenAI, Google, ElevenLabs, Cartesia, Camb, Rime, Sarvam, Speaches and Azure.

Each exposes its own voice list and its own tunable settings. Speech rate is available broadly; expressive controls are provider-specific — Azure's styles and Rime's sampling controls on its arcana model appear only when you pick that provider.

If a provider's parameter is not in the voice settings panel, it is not available from the builder today, whatever the provider itself supports. The panel is the source of truth for what you can control.

Auditioning a voice

Voice Bench has two modes.

• Single renders your script through one configuration, so you can tune settings and hear each change.
• A/B Compare renders the same script through two configurations, so the difference is the only variable.

Phone quality (8kHz) re-renders the preview at the sample rate a phone line actually carries. A voice that shines at full bandwidth can lose its edges there, so judge on the phone-quality preview before committing.

Saving a profile

Once a configuration sounds right, save it as a named Voice Profile. Tick Include Config B as a fallback voice first to store the second A/B candidate as that profile's fallback.

Name profiles after the agent they belong to rather than the voice — "Support line", not "Cartesia Sonic". The provider behind a profile can change later; what the profile is for does not.

Using a profile in a workflow

A profile does nothing until a workflow points at it. Open the workflow's model configuration and pick a Voice Profile; its settings are expanded into the workflow's speech configuration at call time.

Two consequences worth knowing:

• The reference is live. Edit a profile and every workflow pointing at it changes on the next call, with nothing to republish.
• The choice is per workflow, and it applies to every node in that workflow that speaks. There is no per-node voice override.

Expressive Mode sits beside the profile picker. It adds the selected provider's own emotion markup to the model's instructions, so the agent can direct its delivery from what the caller said. The markup language differs per provider and one provider's tags are read aloud as words by another, so the toggle is disabled unless the profile's model supports expressive delivery.

Next steps