With faster-distil-whisper-large-v3 or large-v3, transcribe instruction is ignored (it translates instead)

[ketsapiwiq]

Hi! With faster-distil-whisper-large-v3 (in reality, Systran/faster-distil-whisper-large-v3 is the correct name) or large-v3, transcribe instruction is ignored (it translates to english instead) in the following code:

With medium model, no problem at all for transcribing (input audio is French).

Code:

import speech_recognition as sr

recorder = sr.Recognizer()

from faster_whisper import WhisperModel
audio_model = WhisperModel("Systran/faster-distil-whisper-large-v3") # or large-v3

audio_data = recorder.listen(
            source
        )

audio_np = (
        np.frombuffer(audio_data.get_raw_data(), dtype=np.int16).astype(np.float32)
        / 32768.0
    )
    
result = audio_model.transcribe(
        audio_np,
        language="fr",
        task= "transcribe",

Is it a problem with the model? Should I file this issue elsewhere? It seems it also affects normal Whisper Python lib as well.

[Purfview]

Distil models are not multilingual, it's English only.

[ketsapiwiq]

Ok thank you, I will do more tests with large-v3, it's not systematically translating but only sometimes, if I'm not mistaken.

[ketsapiwiq]

Ok I must have mixed my models up, sorry, closing.