search

ElishaAz / Sayboard

An open-source on-device voice IME (keyboard) for Android using the Vosk library.
GNU General Public License v3.0
303 stars 19 forks source link

Bigger models #38

Open HyperCriSiS opened 8 months ago

HyperCriSiS commented 8 months ago

Hi, Would ne really awesome if you could add support for the bigger models :)

Thanks Hyper

ElishaAz commented 8 months ago

I've just tried the standard English model on a virtual device (mine doesn't have enough RAM)

The current app can actually run the bigger models without any change, you can install the model manually (I just added a Wiki page showing how to do this)

Maybe I'll add an option to disable the verification of the model, to allow importing it within the model

HyperCriSiS commented 8 months ago

OK nice, I will try 😊

HyperCriSiS commented 8 months ago

Unfortunately it keeps crashing. My phone has about 9GB free RAM. I tried the 1.9 GB model. It works for a while and then the RAM usage takes another heavy spike and it crashes. There seems to be some strange bug. while loading it typed some words without me saying anything. Also correction of words was not properly possible. One word was constantly underlined and could not be deleted. Also it is not possible to stay on the back button, every character has to be deleted manually.

The speech recognition itself was outstanding, really incredible how good and fast...

Do you plan to add support for a vosk server?

nshmyrev commented 8 months ago

You can try big model without rnnlm and rescore folders, it will be more accurate than small model and will take about 1Gb of memory, not 16 like currently.

HyperCriSiS commented 8 months ago

Thank you, that works and recognition is still pretty good. Awesome that this can be adjusted this dynamically.

The RAM usage is really crazy. Was not aware of that.

pixel2user commented 8 months ago

@HyperCriSiS, where can I download a "big model without rnnlm and rescore folders"?

HyperCriSiS commented 8 months ago

@HyperCriSiS, where can I download a "big model without rnnlm and rescore folders"?

I downloaded a 1.9GB model for my language. I installed the zip via Sayboard and deleted the two folders afterwards.

LuccoJ commented 7 months ago

@pixel2user presumably you have to remove those two folders from the zip file.

Charles7z commented 3 months ago

I compared the small and large module on my pixel 6a. And after a few tests it was clear that the large module did far worse then the small module in this particular instance, maybe it would be different with a phone with more ram.

However i did not delete the two folders because i could not find them in the zip file. So maybe it would be different if i did it a different way than loading the zip directly via the app. But either way i think the small module works really well, and is very fast on my phone.