Automatic Speech Recognition
Transformers
Safetensors
German
whisper
Eval Results
Inference Endpoints

german or swiss-german

#5
by jschoene - opened

Your german turbo model works very well* in faster-whisper

  • no hallucinations
  • no repetitions
  • no errors in upper/lower case letters
  • very fast, almost as fast as the new OpenAI Turbo Model

but there is a “ß” problem, some examples

  • wrong “Massstab”, correct would be “Maßstab”
  • wrong “Grösse”, correct would be “Größe"
  • ...

I have about 150 such misspelled different words in about 15 hours of video
is it possible that your additional dataset was not german, but swiss*-german?

  • with "condition_on_previous_text" = True :)
    ** in switzerland there is no “ß”, it is replaced by “ss”
primeLine AI Services org

Yeah, I think that can be due to a typo in the preprocessing of the training dataset.
While casting all into Unicode characters and some chars like ß, ü,ö,ä etc back this can occur

a had a look in your used dataset "asr-german-moxed-evals"

  • common_voice_19_0
  • multilingual librispeech
  • Tuda-De

in all three datasets are wrong "ss" chars and parts of the two first are using sometimes (very) outdated german

in the first 35 entries I have found (column "references")

wrong "ss", should be "ß",

  • liess -> ließ
  • verschloss -> verschloß
  • dass -> daß
  • mussten -> mußten
  • dass -> daß
  • Schweiss -> Schweiß
  • Grosse -> Große
  • aussen -> außen
  • äusserlicher -> äußerlich

wrong spelling (missing char) and outdated german

  • wusst -> wußte ('e'missing )
  • müsst -> müßte ('e' missing)
  • offenliess -> offen ließ (neue Rechtschreibung)

correct "ss"

  • müssen
  • Sessel
  • unablässig
  • fallenzulassen
  • umfasste (ok, neue Rechtschreibung)

one would have to be corrected the 9799 entries !!! - since this will probably not happen, a fine-tuned model with this data is unfortunately not really suitable - a real pity

Sign up or log in to comment