Exceptions: java.lang.OutOfMemoryError: Java heap space

[5exceptions-rakeshdiwan]

Unable to integrate "vosk-model-en-us-0.22" language model with assets because of it's large size.

I'm working on one of the requirement of Android application and trying to integrate the large model into application package, I've encountered accuracy issues with small model so I switched to large ones. Please find details below about what I've tried so far:

• Successfully build sample application with small model but facing packaging error(java.lang.OutOfMemoryError: Java heap space) with large models while placing model into assets directory.
• Downloading this model at runtime: It's causing network timeout error from server and unable to fetch it.
• Manually download model and store somewhere in public directory but the model initialization process throughs and OutOfMemoryError. I guess this is because SDK internally trying to unzip the model into private directory.

Please assits me to get more proper way to integrate VOSK into android or flutter to get the more accuracy at runtime.

[nshmyrev]

Big models are created for big servers, not for mobile phones.

To get suggestions on accuracy improvement you need to share sample audio data.

[5exceptions-rakeshdiwan]

Can I share audio and other details here or we can use private email thread for that? Please assist me for this.

[nshmyrev]

You can share here.

[5exceptions-rakeshdiwan]

@nshmyrev Due to maintain confidentiality of work I’m not able to attach audio file here in this chat. Instead of it I’m sharing a drive link for all the required files in a folder. Please request for access on the same or give me email for access.

Folder contains:

1. Sample audio file.
2. Idea and issue brief video.

Links

1. Drive Folder: https://drive.google.com/drive/folders/11jXpjNbqCU0tvk7kUrX47fsqHXkLvP9m?usp=share_link
2. List of words and phrases want to target: https://docs.google.com/spreadsheets/d/1aSjcTkF58qBJiLMCteIh_n_nCj5lz200ayIDY8ZlKZg/edit?usp=sharing

Thanks in advance.

[nshmyrev]

It doesn't allow me to download, says I need to request access.

[5exceptions-rakeshdiwan]

It doesn't allow me to download, says I need to request access.

The link is open can you please try now?

[nshmyrev]

Well, you need better microphone. The current one is awful and cuts audio at 3khz. It has nothing about model size.

[5exceptions-rakeshdiwan]

It has nothing about model size.

Can I have any reference audio file?

[5exceptions-rakeshdiwan]

I've managed to record the phrases again please check and let me know if the cuts and pause time are correct. https://drive.google.com/drive/folders/1xtU0eSp8uf0wEfXHwRyIOw5pvxiCu0H8?usp=sharing

[nshmyrev]

Now the audio quality is much better

[5exceptions-rakeshdiwan]

Great, But I'm still facing the accuracy issue

[5exceptions-rakeshdiwan]

@nshmyrev! Can you please advise us on accuracy and how can we resolve this?

[gvoll]

Happy New Year @nshmyrev! I'm the 5exceptions client :) We'd greatly appreciate your guidance on how to improve the accuracy of these commands to support the voice activation-based app we're working on. Please let us know if there's a different preferred way to provide these audio files or anything else that will help move this to resolution. Ideally, this will be an approach that we can repeat when adding commands going forward.
Thank you!

[nshmyrev]

@gvoll you can bias model to specific commands, see

https://alphacephei.com/vosk/lm

https://github.com/alphacep/vosk-api/blob/master/python/example/colab/vosk-adaptation.ipynb