Wondering how to train small models, big models work

[Binozo]

Hello there 👋

recently I successfully trained a new word for the german vosk-model-de-0.21 :)

As I am currently running vosk on my Pi 5 I have limited memory and that german model already exceedes 5GB in memory usage. I wanted to perform that training on a smaller model like vosk-model-small-de-zamia-0.3) or vosk-model-small-de-0.15 as I already had really good results with those in my case.

On your page is the following instruction: For small model, just pick the required files from exp/chain/tdnn/lgraph. but sadly that lgraph directory is nowhere to find.

I found a reference to that directory in compile-graph.sh:8: rm -rf exp/lgraph but nowhere else 🤔

All other required directories and files for the big models do exist and work as expected. Does anybody know how to compile for those smaller models?

[nshmyrev]

You need a small model compilation package, you can request it over email contact@alphacephei.com

[Binozo]

Ok thanks!