Add doc about FST-based CTC forced alignment.

[csukuangfj]

It is based on CTC FORCED ALIGNMENT API TUTORIAL from torchaudio, but we are using
an FST-based approach.

I can produce identical output with torchaudio using https://github.com/k2-fsa/kaldi-decoder.

I am refactoring the code and will prepare at least two colab notebooks.

[whaozl]

Hi, The align tool can make the word time stamp is accurate on the begin and end postion ?

[csukuangfj]

Hi, The align tool can make the word time stamp is accurate on the begin and end postion ?

It depends on what model you use.

You can have a look at https://pytorch.org/audio/main/tutorials/ctc_forced_alignment_api_tutorial.html
We can produce identical results with torchaudio using the same model.

[lifeiteng]

@csukuangfj Will it be completed soon?

[csukuangfj]

@csukuangfj Will it be completed soon?

Yes. I am working on it now.

[lifeiteng]

@csukuangfj Has k2-based approach been forgot?

[csukuangfj]

No, it is TODO.
Please use the first approach at present or you can add the second approach with k2 by yourself.
All APIs you need are there. You only need to combine them.

[cageyoko]

@csukuangfj Sorry to bother you, and look forward to your reply. Recently I compared several different alignment methods, such as TorchAudio(ctc), whisperX and funasr, and found that none of them were as good as Kaldi-based alignment. The conclusion is consistent with this paper "https://www.isca-archive.org/interspeech_2024/rousso24_interspeech.pdf", do you have some advice or todo in alignment in k2-fas project.

[danpovey]

Kaldi's TDNN systems have limited context, which may give better alignment. (Or GMMs may give even more precise alignments as they have even less context). I'm not sure that alignment is a super big priority in k2-fsa right now, is there a specific type of application you have in mind?

[cageyoko]

Thanks to Daniel for solving a problem I have had for a long time (less context is more important for alignment?). Yes. I have been working in spoken pronunciation scoring, which relies heavily on the accuracy of phoneme or word alignment.