If your first experience with Lexicon wasn’t flawless, don’t fret! There are lots of things we can try to improve.
Cloud Sync Issues
If for some reason the Speech to Text or Assistant sync fails, check the log for any error messages. You may have entered your credentials incorrectly, may have exceeded your plan limits, or may have invalid data.
- The Assistant Lite plan has a limit of 5 workspaces. If you need to delete a workspace to free up space, go to the online tool: Watson Assistant
- If the Speech to Text sync gets stuck in a loop with error messages, cancel and try again in 5 minutes. If it still doesn’t work, unlock the Customization Id in the Workspace Settings, delete the id, and resync.
- If either sync is failing inexplicably, try restarting Unity.
If your speech to text results are garbled, we need to check that your microphone is up to the task. One simple way is to try the Watson Speech to Text online demo: Watson Speech to Text Demo. Try this in a browser that supports microphone input (e.g. Chrome). If you’re getting poor results here, you may want to upgrade to a better microphone.
Speech to Text Accuracy
If you’re using the Lite plan for Watson Speech to Text you may experience some issues with transcription accuracy. For example, a phrase like “create a red cube here” might be difficult for the service to understand (cube is not a common term in everyday conversation). We’ve added the synonym “box” for cube, so you could try “create a red box here” instead. When using the Lite plan, you have to be especially careful not to use domain-specific words. Consider upgrading to the Standard plan and creating a Custom Language Model to improve accuracy.
Lexicon processes the microphone stream to detect speech vs silence. Segments of silent audio are not sent to the cloud. This decreases your Watson usage, but is also important for delineating phrases and lining up timestamps. There is a Silence Threshold setting on the Lexicon Runtime component that you can adjust to ensure silence is correctly detected. To debug, add a SpeechStatusWindow prefab from the Samples/Prefabs/Windows folder. Make sure that the speech status updates correctly while you talk. If not, adjust the Silence Threshold until both speech and silence are correctly detected. If adjusting at runtime, restart the Lexicon Runtime (disable/enable the component) for the changes to take effect.
The Silence Threshold may need to be adjusted for each new microphone and environment. Future versions of Lexicon will dynamically determine this value, but for now you may need to adjust it when you deploy to different hardware or enter noisier/quieter environments.
Lexicon does its best to align each spoken word with pointer (e.g. mouse or gaze) input. You may find that this alignment feels a little off. There is a Timestamp Offset setting on the Lexicon Runtime component that you can adjust to tweak the alignment. To debug, add a DebugWordAlignment prefab from the Samples/Prefabs/Debug folder. This will render each word in 3d space as it is spoken. Try changing the Timestamp Offset setting 0.05 seconds at a time, until phrases like “move this one over here” work correctly. If adjusting at runtime, restart the Lexicon Runtime (disable/enable the component) for the changes to take effect.
The Timestamp Offset may need to be adjusted for each new device and each user.