From 8dd6a9c0aad0b22a0f055746f9720a3d3ba605eb Mon Sep 17 00:00:00 2001 From: Nicholas Riley Date: Sat, 7 Sep 2024 18:48:01 -0400 Subject: [PATCH] Document a common issue reported on the Talon Slack. --- docs/Quickstart/improving_recognition_accuracy.md | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/docs/Quickstart/improving_recognition_accuracy.md b/docs/Quickstart/improving_recognition_accuracy.md index 73f48097..847aeaf0 100644 --- a/docs/Quickstart/improving_recognition_accuracy.md +++ b/docs/Quickstart/improving_recognition_accuracy.md @@ -54,6 +54,14 @@ Talon lets you configure how long it will wait after you stop speaking before tr ``` Start with the value of 0.4, just slightly higher than the default, and increase only if necessary. +## Talon responds inconsistently or with a delay + +Talon only processes your speech after its voice activity detector (VAD) detects silence of the length specified by `speech.timeout` above. + +This issue is sometimes caused by background noise or non-speech sounds (e.g. breathing) being picked up by your microphone. To confirm this is happening, ensure Speech Recognition > Save Recordings from the Talon menu is checked, then reproduce the problem. Each utterance is saved to a separate FLAC file in the `recordings` folder inside the Talon home folder (Scripting > Open ~/.talon from the Talon menu). If the recordings seem longer than you expect, listen to them carefully or open them in an audio editor to see what is being recorded other than your voice. + +If recordings appear to be the correct length, Talon may be stuck due to a long-running voice command or callback. Look in the [Talon log](troubleshooting#check-the-talon-logs) (Scripting > View Log from the Talon menu) for lines containing `[watchdog]` and `(stalled)`. If these don't make any sense to you, share your log on the Talon Slack. + ## Help improving the Talon recognition engine This will not help you today, but instead help to improve the talon engine for all of us in the future. You can do this by providing speech (and noise) samples: