Voice fallback

Keep the agent talking when a voice provider fails by listing fallback voices on the agent and a workspace fallback voice.

If the voice provider errors during a call, the agent would go silent. Fallback voices prevent that: when the primary voice fails, the runtime switches to the next voice in a list. Put voices from a different provider in the list so one provider outage does not take down all of them.

Agent fallback voices#

Set them under Security & Fallback Settings, Fallback Voices. Use + Add voice to append a voice and the × on a voice to remove it. Voices are tried in the order shown.

Security & Fallback Settings with Fallback Voices

The list is stored in the agent's fallback_voice_ids:

curl -X PATCH https://api.tried.com.au/update-agent/agent_3f1c9a7b2d4e6f8091a2b3c4 \
  -H "Authorization: Bearer $TRIED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "voice_id": "11labs-Rachel", "fallback_voice_ids": ["deepgram-Thalia", "11labs-Sarah"] }'

Field	Type	Default	Description
`fallback_voice_ids`	string array	`[]`	Voices from the built-in catalogue, tried in order when the current voice fails.

Workspace fallback voice#

Under Settings, Reliability, Provider fallbacks, Fallback voice sets one voice for the whole workspace. It is added to the end of every agent's fallback chain, unless it is already the agent's voice or already in its list.

The order on a call is:

The agent's voice_id.
Each voice in fallback_voice_ids, in order.
The workspace fallback voice.

See Reliability for the other provider fallbacks (LLM and speech recognition).

How switching works#

The switch happens when a TTS request for the current voice fails.
Fallback voices use the agent's voice tier, Premium or Standard. Agents on another voice_model (set through the API) fall back on the Standard tier.
If every voice in the chain fails, the call ends with disconnection_reason: "error_tts".

Limits#

Fallbacks are resolved from the built-in voices only. Cloned voices cannot be fallback voices, and the dashboard lists the Tried voices.
The fallback voice sounds different from the primary. Pick voices with a similar gender and accent so a mid-call switch is less noticeable.
Fallback applies to the cascading pipeline. Agents on a speech-to-speech model do not use TTS fallback.

Recommended set-up#

Primary	Fallback 1	Fallback 2
`11labs-Adrian` (Adrian)	`11labs-Drew` (Drew)	`11labs-Dave` (Dave)
`11labs-Rachel` (Rachel)	`11labs-Sarah` (Sarah)	`11labs-Matilda` (Matilda)

Needs keys Fallback needs the platform's voice key, like the primary voice.