Skip to main content
Speech caching reuses generated audio for repeated synthesis requests. It is separate from the uploaded-recording library and from recordings of completed calls. The current workflow setting supports MiniMax TTS; it does not cache native realtime speech-to-speech output.

​
Enable caching

  1. Select a supported pipeline model configuration with MiniMax as its TTS service. See Models.
  2. Open the agent’s Settings → General → Speech Caching.
  3. Enable Enable Speech Caching and choose Save General Settings.
  4. Test repeated phrases with unchanged voice/model settings, inspect the cache metadata, and listen to the generated speech. Publish the draft when ready.
Cached audio has a 24-hour expiry policy and is not a permanent recording. Matching includes the provider request, voice/model settings, output format, and credential identity, within the organization. Similar wording or different synthesis settings need not produce a hit. Expiry, eviction, or invalidation can require synthesis again.

​
Inspect the cache

Open Recordings → TTS cache. Inspect the phrase preview, provider, model, voice, duration, usage count, and last-used time. Search or sort the list and preview an entry when needed. Preview listening does not increment its synthesis hit count. Authenticated API access is available under /api/v1/tts-cache: The list supports offset and limit (1–100, default 25). Use the returned entry ID; it is not a workflow-recording ID. Access uses the selected organization’s authenticated session, not a public embed token.

​
Invalidate or disable

Remove an entry to stop reusing that cached audio, or clear the organization’s cache when appropriate. These actions affect other agents in the same organization that share matching synthesis requests. They do not delete uploaded recordings or completed-call media. A later synthesis can repopulate the cache. To stop a particular workflow using the feature, turn off Enable Speech Caching, save, and publish. Disabling a workflow and clearing stored entries are separate actions. If the list is empty, check the workflow toggle, MiniMax TTS selection, repeated input/settings, expiry, and organization. 503 indicates the cache is temporarily unavailable; do not interpret it as proof that a voice credential is invalid.