Nice model!
Hey @flozi00 ,
really nice model! Congrats! Have you also already have a look at Moonshine Models https://huggingface.co/UsefulSensors/moonshine-streaming-small ?
They are pretty small models, run fully on CPU, have latency < 250ms and pretty good WER (Top 20 in https://huggingface.co/spaces/hf-audio/open_asr_leaderboard ) - but so far no German adaptation. Ideally they would be bi-lingual: German & English.
We're working on it here:
https://github.com/moonshine-ai/moonshine/issues/141
Best
fosple
I will take a look how hard it would be to adapt the training script
Thats super nice to hear @flozi00 . I already started adjusting https://github.com/pierre-cheneau/finetune-moonshine-asr (which to the v1 moonshine models) to the new moonshine-streaming v2 models. The code is not yet polished, but if you want to have a look: https://github.com/fosple/finetune-moonshine-asr I tried to simplify the codebase as much as possible so it's easier to read.