Transcription/translation language change -- '50358' / '50359' is not a valid task (accepted tasks: transcribe, translate)

[grzegorz700]

Bug scenario

In the following scenario:

1. model = whisperx.load_model(..., language='es') (model e.g. large-v2)
2. model.transcribe(..., language='es')
3. model.align(..., language='es)
4. model.transcribe(..., language='en')
5. model.align(..., language='en')

We got the error names:

ValueError: '50359' is not a valid task (accepted tasks: transcribe, translate)

Cause

It's based on the sub-case of one problem (default params etc.) In the following lines of code

whisperX/whisperx/asr.py

Lines 201 to 205 in f2da2f8

	task = task or self.tokenizer.task
	if task != self.tokenizer.task or language != self.tokenizer.language_code:
	self.tokenizer = faster_whisper.tokenizer.Tokenizer(self.model.hf_tokenizer,
	self.model.model.is_multilingual, task=task,
	language=language)

So, for the default method param (task = None), it gets tokenizer.task. But before that, the task is mapped to int by faster_whisper.tokenizer.Tokenizer. In consequence, it gets an int name called 50359 (transcribe=50358, translate =50359).

The problem exists when we change the tokenizer language for the existing model wrapper.

Temporal fix, use following:

model.transcribe(...,  task="transcribe")
# or 
model.transcribe(...,  task="translate")

Desired fix:

To reverse map the id of the task from the tokenizer.

Other info about that problem:

#huggingface/transformers#22331