Voice (TTS) configuration

The Voice Configuration section controls the text-to-speech step. It is split into three sub-sections:

1. Voice Selection — provider, voice, custom voice ID, preview.
2. Voice Settings — fine-grained voice characteristics (provider-dependent).
3. Fallback Voices — backup providers if the primary voice fails.

Voice Selection

Provider and voice

Pick a TTS provider — typically ElevenLabs, Deepgram, OpenAI, or Cartesia — and then a voice from that provider's list.

The API Key selector appears below the provider for BYOK. See BYOK.

Custom voice ID (ElevenLabs)

When the provider is ElevenLabs, you can type a custom voice ID in addition to selecting from the pre-populated list. This represents a voice you've created in your own ElevenLabs account (cloned voice, fine-tuned voice, etc.).

• The field accepts free-text input.
• The value is saved as part of this agent's configuration. It is not stored as a reusable resource — to use the same voice in another agent, enter the ID again there.
• If at runtime the custom voice ID errors out, the first voice for this provider is used as a runtime fallback so the call doesn't fail. The fallback is recorded in the trace.

Voice preview

Below the voice dropdown, a Preview voice control plays a short sample statement spoken in the selected voice.

• The sample text is shown next to the play button so you know what will be spoken.
• The sample is pre-defined per voice (or a platform default if the provider doesn't supply one).
• When a custom voice ID is entered, the preview button generates audio using that ID. If the ID is invalid, an inline error is shown.

Voice Settings

These fields tune the characteristics of the synthesized voice. Availability depends on the provider — the UI shows only the fields your selected provider supports.

Setting	Range	Default	Helper text
Speed	0.8 – 1.2	1.0	Slower ↔ Faster.
Stability	0.0 – 1.0	0.5	Higher values produce more consistent output. Lower values add variation.
Similarity Boost	0.0 – 1.0	0.75	How closely the voice should match the original voice profile.
Style	0.0 – 1.0	0.0	Exaggerates the style of the voice. Higher values are more expressive.
Use Speaker Boost	toggle	Off	Enhances speaker clarity at the cost of higher latency.

Provider-specific availability

• ElevenLabs — all settings available.
• Deepgram, OpenAI, Cartesia — only Speed is available; the others are hidden.

When you change provider, values for unsupported settings are discarded — no warning is shown, the fields simply disappear.

Fallback Voices

A collapsible Fallback Voices section lets you add backup TTS providers for resilience.

• If no fallbacks are configured: No fallbacks configured. Add a fallback provider for improved reliability.
• Click + Add Voice Fallback to add a row with provider + voice selectors and a trash icon.
• You can add up to 3 fallback voices.

How fallback runs at runtime

• A failure is an exception or a timeout from the primary TTS provider.
• Fallbacks are tried top-to-bottom in the list.
• No fallback after the audio threshold: once audio playback for a turn has started and exceeded 3 seconds, the platform will not switch to a fallback — it completes with the current provider or fails gracefully. This is deliberate, to prevent jarring mid-sentence voice changes.
• When a fallback was used for a turn, the TTS span in the trace carries a provider_fallback_used: true attribute.

Where TTS shows up in the trace

The TTS span in the run trace includes the agent's response text that was synthesized, plus the TTS generation wait time. See Run history & traces.