Hi there,
i just setup a local language pipeline with rhasspy faster whisper and would like to know if there is any better language model to try in german than tiny int8.
The recognition rate really is not good in my eyes, doesnt matter if i use the smartphone of the M5 Stack
I installed fast whisper from rhasspy on my Mac Mini M4 and use the medium-int8 model, which works satisfying with German (finally) The tiny model for German is not usable in my opinion. It was a very frustrating experience with the voice feature, but now it feels good.
Wondering if there is no guessing or similarity routine implemented.
I get frequently barley wrong words like “an” is understood as “am” so it does not work at all.
For example, the words are “Weihnachtsbaum” and “kleine Lampe”.
INFO:wyoming_faster_whisper.handler: Weihnachtsbaum, Pann!
INFO:wyoming_faster_whisper.handler: Weihnachtsbaum an.
INFO:wyoming_faster_whisper.handler: Kleine Nampel an.
INFO:wyoming_faster_whisper.handler: Kleine Lampere an
There we get easily 50%+ false detects (at tiny and medium with beamsize 5) because it looks like whisper does not know what words are actual available. I tested 1-2 years ago a setup with prelearned words at garbage hardware and got basically 100% success rate.
What device are you using for voice recognition? Right now I am using ReSpeaker Lite. It works ok so far when I speak loud and “clearly”. But its annoying to concentrated on the spoken. I would like to talk on my sloppy way, like I did when I had Alxa in use.
I have much more hopes in the Home Assistant Voice Preview Edition.
UPDATE: I found this German model which I will test now
For now it works good, but a bit to large for my Mac Mini. It not snappy enough compared to the medium model
Update 2:
Right now I switched to Whsiper.cpp
I installed it on my (Mac Mini M4) LLM Server with Metal support. I have the large-v3 model running. Thanks to Metal I have instant response even with the large model.
To get whisper.cpp intergrated into wyoming I have to use following