Advanced FluidAudio Settings

Find this section in Speech Engine → FluidAudio → Advanced FluidAudio.

Voice Trigger

Voice Trigger controls what the engine treats as speech before phrase matching starts.

• Very Strict: least permissive, lower sensitivity to background sounds
• Strict: more filtering of marginal speech
• Default: balanced baseline behavior
• Permissive: accepts quieter or partial speech with fewer pauses
• Very Permissive: most permissive to speech activity

Internal mapping:

• Default is the preset value (0.65), which is the app default.
• The slider is quantized to the five presets above.
• The value is shown as a VAD threshold and appears in parentheses for visibility.

Footer behavior: this setting is specifically about speech activity detection, not model selection.

Reset option:

• Reset to Defaults restores the default trigger preset.
• Disabled automatically when the current setting is already at default.

Delete Cached Diarization Models

These actions remove downloaded model caches only; they do not delete your profile data, phrases, or language settings.

• Delete Sortformer Models
• Delete SpeakerManager Models
• Delete LS-EEND Models

What happens when you delete:

• Any active warm-up state for that backend is cleared.
• The next session will re-download required model files for that backend.
• The action is blocked when no cache exists for that backend.

NOTE: Model downloads can take a while depending on network speed, so delete models only when necessary. After deletion, the next selected backend session will re-download files before it can run.

Best Practices

Use Voice Trigger to balance false starts vs missed speech:

• Choose Very Strict in loud or command-heavy environments with lots of noise.
• Choose Default for most normal use.
• Choose Permissive/Very Permissive when users speak softly, at distance, or in acoustically weak conditions.

Manage model cache only when needed:

• Use delete actions only if models feel stale, you are troubleshooting backend startup issues, or you need a clean refresh.
• If changing between Sortformer and LS-EEND frequently, this can be helpful when trying to force a fresh model state on each backend.