Wake Word detection - Githubissues

adhitht commented 2 years ago

I have downloaded g-assist from snap store version 1.0.0-rc.2. Wake word doesn't seem to work. (I am not sure if it is built for the version) Please clarify. Saw that wake word detection is built using bumble-bee. I have tried bumble-bee from here : https://jaxcore.github.io/bumblebee-hotword/full-example/. I have tried hey_google and ok_google. But bumble-bee doesn't seem to recognise both. (Had 1 recognition with almost 20 tries :disappointed: ). Alexa seems to hear me. It would to be better if you could just provide option to set other wake-words in settings. I would also suggest teachable machine from google as an alternative to bumble-bee. (It works very well.) Hope you could implement that well. Teachable machine can be customised to user so recognition will be far more effective. https://teachablemachine.withgoogle.com/ snowboy is another depreciated software which can be used to have wake-word detection.

im-coder-lg commented 2 years ago

Build from the codebase via NPM, and the wakeword will work. Since the new release isn't out yet(but will be soon) you will have to build it from source. Get Git and Node JS 16, clone the repo in a folder(random) and run the commands listed in the Readme. Then, open the newly-created dist folder and you might have to check that. Run the installer there(you bult it so you are secure) or check the linux-unpacked(doubt here but unpacked confirmed) folder and run the Google Assistant executable for Linux. Then, you are good to go.

im-coder-lg commented 2 years ago

Teachable machine can be customised to user so recognition will be far more effective. https://teachablemachine.withgoogle.com/

Wow, @Melvin-Abraham this is a good idea, right? Bumblebee currently can't recognise Ok Google or Hey Google, which means you have to release a little(in my case, a lot) sound, but if we try integrating this, it might recognise Hey Google/Ok Google perfectly, right?

im-coder-lg commented 2 years ago

Bumblebee currently can't recognize Ok Google or Hey Google,

Sorry for that... It was actually: Bumblebee currently can't recognize Ok Google/Hey Google quickly or in one sentence, ie. Ok Google.

adhitht commented 2 years ago

Bumblebee currently can't recognize Ok Google/Hey Google quickly or in one sentence,

Well keyboard shortcut would be better than that.

Melvin-Abraham commented 2 years ago

I tried some other wake words in bumblebee which seem to work much better. Maybe these could be provided as possible wake-word options...

Melvin-Abraham commented 2 years ago

Teachable machine can be customised to user so recognition will be far more effective. https://teachablemachine.withgoogle.com/

Wow, @Melvin-Abraham this is a good idea, right? Bumblebee currently can't recognise Ok Google or Hey Google, which means you have to release a little(in my case, a lot) sound, but if we try integrating this, it might recognise Hey Google/Ok Google perfectly, right?

This would require a ton of voice data from a number of people with different accents and dialects. If it is trained with a small sample group, it would perform worse for other set of people...

So, my plan for now is to keep using bumblebee with extended functionalities for this release. If that doesn't suffice, maybe we can give it a thought

adhitht commented 2 years ago

You can download dataset for one person,so it can be stored at a particular location and accessed by the app. So it will work perfectly for one user.

im-coder-lg commented 2 years ago

hey, why not keep one audio file for the voice recog(recognition)? We can use the same way Androids ask for Ok Google and Hey Google twice, right? Can that work?

adhitht commented 2 years ago

Did not get what you are talking about. Let me clarify, with teachable machine you have to record background voice and a class, as we are intended to make a wake word, we have to repeatedly say something like Ok Google, or Hey Google, (or both together) and it should be minimum 8 samples for class. At last we have to train model and export data, it gives some zip file with almost a few kilobites of data (two json files and a bin file). We could use them to trigger wake word. So the audio samples have to be collected by user and add it in a directory and make g-assist read that. (Same as in the case of tokens and credentials file). It would be awesome if you could implement that.

Melvin-Abraham commented 2 years ago

I am not sure if teachable machine is a perfect candidate for this use-case. If I am not wrong, it is best fit for distinguishing different quality/types of sounds, different timbre, tones, etc. but not fit for a wake-word detection as it could lead to a plethora of false-positives or false-negatives which could possibly trigger the assistant every-so-often if not trained properly. Also, that might require more than 500 to 1000+ (not a calculated guess) samples for both random noise (false triggers would mostly depend on this) and wake-word classes just for a single person for it to even perform properly which would require a lot of training, thus, space and time overhead. That's just my opinion though...

im-coder-lg commented 2 years ago

I say copying the Android Google Assistant Voice Detection setup might work. Two Ok Google audio files and two Hey Google Audio files or just go get them from the global Google Account since they are stored on the Google Account.