Results 1 to 5 of 5

Thread: Trigger word recognition and vr overlay offsets not fully saving

  1. #1
    Junior Member
    Join Date
    Sep 2020
    Posts
    4

    Trigger word recognition and vr overlay offsets not fully saving

    I've got two issues as mentioned in the title

    It seems that the voice recognition is very confident that i've said the trigger word every time it hears a sound. Just a yawn sets it off and its confident that i've said it. Tried changing the word but that doesn't seem to make a difference. CC otherwise has no problem understanding anything i ask it.

    The other issue is that my z-offset is not saving for overlays in vr. I'm using an electron overlay to host the kapps relatives window and using cc to bring that overlay into vr. I can move it around and get it in place with ease, and the x/y and rotation will all save, but the z offset goes back to the default of -1.0. EDIT: The z offset is reset after closing CC. It stays in place after i've set it, but when i shut the app down and restart it the z changes back.

    I'm on a fresh install of windows and CC as of yesterday, the voice recognition was happening both before and after the fresh install, the vr offset issues is new as of the reinstall. Any help would be appreciated, thanks.
    Last edited by Cptnspoon; 23-09-2020 at 19:02.

  2. #2
    Crew Chief Mega Corp CEO mr_belowski's Avatar
    Join Date
    Feb 2017
    Posts
    1,815
    will have to wait for Morten to offer some insight into the VR overlay issue, but i'll take a look at the trigger word stuff tomorrow morning. I don't use it (I prefer hold-button) so it doesn't get much internal testing

  3. #3
    Crew Chief Mega Corp CEO mr_belowski's Avatar
    Join Date
    Feb 2017
    Posts
    1,815
    quick update on this


    I think Morten has fixed the VR overlay issue.

    Looking at the trigger word stuff, the issue here is that the confidence values for the built-in Windows speech recogniser are quite different to the corresponding confidence values for the Microsoft recogniser. The Windows recogniser (the one you train) tends to show high levels of recognition confidence (generally between 0.5 and 0.9 for commands and 0.9 and 0.99 for the single word trigger recognition). The Microsoft recogniser (the one you have to install separately) tends to show lower levels of confidence (generally between 0.2 and 0.8). The app only has one set of 'minimum confidence' properties and their default values are appropriate for the Microsoft recogniser and so are much too low for the Windows recogniser. This makes the app too sensitive to matches when using the Windows recogniser. This isn't a major issue with regular voice commands because the system will offer its best guess as to the recognised text (so we're not relying on confidence values so much) but the when recognising the trigger word we need to be pretty sure as false-positives are a big problem.

    To address this I've split the confidence values into 2 sets (one for each recogniser). The Properties form will only show the values appropriate for the recogniser you've got selected when starting the app, so if you start the app with the Windows recogniser the default values for "Minimum voice recognition confidence", "Minimum name voice recognition confidence" and "Minimum trigger word confidence" will be much higher (0.9 or more, I'm still finding appropriate values). So with the next release the trigger stuff should be far less prone to false positives when using the Windows recogniser and this change won't require users to tweak the app.

    I'll get this tested and added to the next release. In the mean time, you could experiment with the confidence values for "Minimum voice recognition confidence", "Minimum name voice recognition confidence" and "Minimum trigger word confidence" - maybe 0.95 for the trigger one would work?

  4. #4
    Junior Member
    Join Date
    Sep 2020
    Posts
    4
    Awesome thank you, I will give those a shot this evening and let you know if they fix the issue.

  5. #5
    Junior Member
    Join Date
    Sep 2020
    Posts
    4
    So I remembered that i had not gone through the speech recognition training after reinstalling windows(I had done it before when i first started having the issues before the reinstall), so I gave that a shot to see if it would help. This is the result of two yawns followed by me actually saying the trigger word(in bold).

    Code:
    17:26:39.481 : Heard keyword Chief, waiting for command confidence 0.9845313
    17:26:39.532 : Opened channel - waiting for speech
    17:26:39.532 : Getting audio from default device
    17:26:44.850 : Gave up waiting for voice command, now waiting for trigger word Chief
    17:26:47.946 : Heard keyword Chief, waiting for command confidence 0.9871724
    17:26:47.998 : Opened channel - waiting for speech
    17:26:47.998 : Getting audio from default device
    17:26:53.333 : Gave up waiting for voice command, now waiting for trigger word Chief
    17:26:57.280 : Heard keyword Chief, waiting for command confidence 0.9899238
    17:26:57.331 : Opened channel - waiting for speech
    I then restarted CC and started getting better results. Not sure if this is related, but it seems if you do the speech training while cc is running you need to restart cc for it to take effect. Yawns are no longer setting it off, however any word that resembles chief, such as see, or cheese, etc, still occasionally set it off, but i assume thats to be expected when chief is throwing 0.95 and cheese is 0.949 theres likely to be some overlap there. I've dropped from 0.98 recognition for most sounds to .89-.94 for words resembling the trigger word. I set the trigger word confidence to 0.95 and that seems to be helping, as long as i don't get into a conversation about cheese while i'm racing. Thanks for the help

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •