Speech recognition settings

Setup

Plugin pages are hidden from the sidebar until you turn them on; see showing a plugin's pages.

Open SpeechToText in the sidebar.

  1. Under Quick Settings, leave Use default microphone on, or turn it off and pick a Microphone Input. The Live Microphone Input meter shows whether it hears you.
  2. Pick a Whisper Model. The default, TinyEn, is small and fast and understands English. Larger models, such as BaseEn or SmallEn, are more accurate but slower. Unknown names fall back to TinyEn.
  3. Click Save Quick Settings, then Start.

The first time you start with a model, MultiShock downloads it, which can take a while. To start listening automatically, turn on Auto-start listening on app startup. The triggers below only fire while it is listening.

Recognition works on chunks of audio. Segment (ms) sets how long each chunk is (3.5 seconds by default); longer chunks catch whole sentences but take longer to react. Chunks quieter than Silence Threshold are skipped, and results below Min Confidence are ignored. The page lists recent sentences it recognized, which helps when tuning these.

More options, such as the recognition language (Recognition Culture, or auto to detect it) and a custom model file, are under Manage plugins → Configure on SpeechToText.

Language and model files

Under Manage plugins, open Configure for SpeechToText. Set Recognition Culture to the intended language or auto. Whisper Model Path (optional) lets you choose a local model file; Auto-download model when missing controls whether the plugin downloads a missing model.

Stop listening before changing the microphone or model, save the settings, then start again. Check the page's status if startup fails or a model is unavailable.

Tune recognition

Start with the microphone meter: a model or word filter cannot fix an input that is not receiving audio. Then compare the recent sentences with what you said.

  • If quiet speech is skipped, review Silence Threshold and microphone input level.
  • If phrases are cut between chunks, increase Segment (ms), accepting a longer wait before the result arrives.
  • If noisy recognition reaches flows, tune the recognition-level confidence threshold and then the individual node's Min Confidence.

Auto-restart listener after unexpected stop and Restart Delay (ms) are available in the plugin configuration. They handle unexpected stops; use Auto-start listening on app startup for launching with MultiShock.

Configure word and sentence triggers

Use Sentence Recognized to match a complete recognized sentence, or Word Recognized for individual words. Set a filter and Match Mode on the node: Any ignores the filter, Contains accepts a partial match, and Exact requires the complete sentence or word for that node. Choose Case Sensitive only if needed.

Recognition and the flow must both be running. If sentences appear in the plugin but the flow does not fire, check the node filter, confidence threshold, and flow running state.

Contributors

Thank you to everyone who has helped improve this documentation.

bwmpbwmp