xinjli / alqalign Goto Github PK

Thank you very much for your excellent work with this tool.

I am currently trying to use it in ESPnet to generate the alignments for training a model with durations similar to the MFA tool.
However, I am facing some minor issues.

I understand the tool is still under construction, but a little guide will help me continue with my implementation.

• you need to add a __init__.py at the alqlign folder so you execute the command alqalign.run.
• In run.py, what is step 1 supposed to do? Transcribe the audio into word/text or just only into phonemes; should it have a similar behavior as step 2?
• In the case of using step 1, text is no longer required, right? So the argument text becomes unneeded? Is it also necessary to move the text processing after step 1 (

  alqalign/alqalign/run.py

  Lines 70 to 92 in 10081d5

  	if text_file.is_dir():
  	for text_path in sorted(text_file.glob('*')):
  	utt_id = text_path.stem
  	if utt_id in utt2audio:
  	audio_files.append(utt2audio[utt_id])
  	text_files.append(text_path)
  	output_dirs.append(output_dir / utt_id)
  	utt_ids.append(utt_id)
  	else:
  	for i, line in enumerate(open(text_file, 'r')):
  	if text_format == 'kaldi':
  	fields = line.strip().split()
  	utt_id = fields[0]
  	sent = ' '.join(fields[1:])
  	else:
  	utt_id = str(i)
  	sent = line
  	if utt_id in utt2audio:
  	audio_files.append(utt2audio[utt_id])
  	text_files.append(sent)
  	output_dirs.append(output_dir / utt_id)
  	utt_ids.append(utt_id)

  ) >>>

  alqalign/alqalign/run.py

  Line 107 in 10081d5
• when using a scp file, do you need to use a text format to load the file?

  alqalign/alqalign/run.py

  Lines 60 to 62 in 10081d5

  	for line in open(audio_file):
  	utt_id, ark_key = line.strip().split()
  	utt2audio[utt_id] = ark_key

  , Is it not possible to just use kaldiio.load_scp to load the file?