Cache folders not created when transcribing audio

[exengo]

I use UltraSinger with a model to transcribe Swedish songs, specifically like this:

python UltraSinger.py -i 'some-swedish-song' --whisper_align_model 'KBLab/wav2vec2-large-voxrex-swedish'

It fails with a FileNotFoundError.
Traceback (most recent call last): File "/home/user/Dev/UltraSinger/src/UltraSinger.py", line 693, in <module> main(sys.argv[1:]) File "/home/user/Dev/UltraSinger/src/UltraSinger.py", line 573, in main run() File "/home/user/Dev/UltraSinger/src/UltraSinger.py", line 147, in run TranscribeAudio(process_data) File "/home/user/Dev/UltraSinger/src/UltraSinger.py", line 353, in TranscribeAudio transcription_result = transcribe_audio(process_data.process_data_paths.cache_folder_path, File "/home/user/Dev/UltraSinger/src/UltraSinger.py", line 483, in transcribe_audio with open(transcription_path, "w", encoding=FILE_ENCODING) as file: FileNotFoundError: [Errno 2] No such file or directory: '/home/user/Dev/UltraSinger/src/output/Kent - Ingen kunde röra oss/cache/whisper_large-v3_cuda_KBLab/wav2vec2-large-voxrex-swedish_KBLab/wav2vec2-large-voxrex-swedish_16_None_None.json'

I figured out that the folders were not created and open therefore failed. Adding this on line 483 fixes the issue:
os.makedirs(os.path.dirname(transcription_path), exist_ok=True)

[Calamdor]

Thanks for this, was trying to figure out why every model I tried was not working!

[agwosdz]

Weird, folder creation occurs in procedure CreateProcessAudio, line 142 - " process_data.process_data_paths.processing_audio_path = CreateProcessAudio(process_data)", as defined in CreateProcessAudio - os_helper.create_folder(process_data.process_data_paths.cache_folder_path) , line (423).

I wonder what is happening there. transcription_path is just a .json file in the cache_folder_path (transcription_path = os.path.join(cache_folder_path, f"{transcription_config}.json"))

Was there any other error message prior?

[rakuri255]

Can someone make an PR or give an song link?

[agwosdz]

Can someone make an PR or give an song link?

https://www.youtube.com/watch?v=17HIRea5C6Y

[agwosdz]

I am pretty sure I know what the problem is. Will create a PR.

Issue is the "/" in the --whisper_align_model option/parameter.

The "/" get's interpreted literally, confusing python about the cache path.

[agwosdz]

Workaround, if you are in a hurry is to edit UltraSinger.py at line 466:

def transcribe_audio(cache_folder_path: str, processing_audio_path: str) -> TranscriptionResult:
"""Transcribe audio with AI"""
transcription_result = None
### whisper_align_model_string = None
if settings.transcriber == "whisper":
### if not settings.whisper_align_model is None: whisper_align_model_string = settings.whisper_align_model.replace("/", "_")
### transcription_config = f"{settings.transcriber}{settings.whisper_model.value}{settings.pytorch_device}{whisper_align_model_string}{settings.whisper_batch_size}{settings.whisper_compute_type}{settings.language}"
transcription_path = os.path.join(cache_folder_path, f"{transcription_config}.json")
cached_transcription_available = check_file_exists(transcription_path)
if settings.skip_cache_transcription or not cached_transcription_available:
transcription_result = transcribe_with_whisper(
processing_audio_path,
settings.whisper_model,
settings.pytorch_device,
settings.whisper_align_model,
settings.whisper_batch_size,
settings.whisper_compute_type,
settings.language,
)
with open(transcription_path, "w", encoding=FILE_ENCODING) as file:
file.write(transcription_result.to_json())
else:
print(f"{ULTRASINGER_HEAD} {green_highlighted('cache')} reusing cached transcribed data")
with open(transcription_path) as file:
json = file.read()
transcription_result = TranscriptionResult.from_json(json)
else:
raise NotImplementedError
return transcription_result

Marked and highlighted are the changes:

Essentially, we check if the argument for --whisper_align_model was provided, and if it was, we replace any "/" in the name with "_" in the cache folder path. That way, the setting remains valid for the model, but does not confuse the OS with a "/" in a pathname, making it interpret it as a folder and hence resulting in a file not found error.