[workflow] Word-level forced alignment with pretrained models from Torchaudio

[pzelasko]

As a follow-up to #824 I'm adding another workflow for forced word-level alignment. It may be interesting to combine the two workflows together. In the end I used the torchaudio model because the tutorial example looks quite convincing with regard to word timestampt accuracy, but I did not evaluate if it's better or worse than any other approach. We can add other workflows for forced alignment later to create some choice.

@desh2608 I also refactored the AlignmentItem thing a bit so that it's NamedTuple again, for the sake of efficiency. I don't like this approach though, after using it for some time. I think we should eventually change it to something like a numpy array (or a collection of them) and maybe not store them in the manifest. Anyway, at least the size of the manifests is greatly reduced with the NamedTuple approach.

[desh2608]

Super-fast development on this one! LGTM, in general.

Regarding alignments, I agree it's not very efficient. Perhaps treating them in the same way as Features would be a better option.

[desh2608]

Here's one approach off the top of my head. We can have an AlignmentObject that contains 2 things:

1. A dict which is just a mapping from a symbol (e.g., word, phone) to an int.
2. A path to a numpy file, containing rows of the form (symbol_idx, start_time, end_time, score).

[csukuangfj]

What does the score indicate?

[pzelasko]

It's aligner-specific "confidence score", in case of torchaudio they suggested to provide the average token probability for a given segment in their tutorial; but it could be also computed differently. I'll add a proper note.