Bad whisper speech recognition ( german )

Hi guys,
found your project and find it absolute interesting.
Running an ryzen ai max+ 395 for my local ki tasks and trying to go completly local for my home assistant voice assistants.
For now i am using ollama with qwen3-vl:30b-a3b-instruct as agent, piper for tts and home assistant nabu casa for stt ( as i know they are also using whisper ). This combination runs absolut perfect.

The last piece for going complete local is a local whisper instance for stt so i gave your solution a try.

It works, but the speech recognition is not very good. Trying to "schalte das Licht in der Küche aus" ends up with "Scheiße das Licht in der Küche aus". It understands "scheiße" all the time wen i said "schalte" That is very bad for a smart home in germany :-)

Starting server with "C:\Windows\System32\cmd.exe /K "C:\Program Files\flm\flm.exe" serve -a 1 --host 0.0.0.0"
HA connects in open ai mode.

configuration.yaml:
stt:

• platform: openai_stt
  api_key: 1234
  realtime: false
  api_url: http://192.168.178.182:52625/v1
  #model: whisper-v3
  prompt: ""
  temperature: 0
  noise_reduction: null

Is this a problem with the whisper npu model?

Start by reproducing the German phrase using the server command in the issue and the configuration.yaml OpenAI STT settings. Compare the resulting transcription with the reported “schalte”/“Scheiße” substitution and determine whether the behavior is specific to the Whisper NPU model; done means the model cause is established or ruled out with supporting output.