Adds metrics "time_span_accuracy", "time_span_precision", "time_span_recall" and "time_span_f1" via pyannote, see detection metrics in https://pyannote.github.io/pyannote-metrics/reference.html
This library seems to be widely used in voice activity detection and speaker diarization. (But it adds a bit many dependencies for my taste. 🙃 Could be made optional maybe.)

Naming of the metrics could maybe be improved? I avoided "coverage" because there is segmentation metric in pyannote with that name.

But these metrics evaluate how much of reference subtitles duration is "covered". For example, recall is the total duration where both hypothesis and reference have a subtitle, divided by the total duration of reference subtitles.

pyannote also has segmentation recall/precision, which - applied to subtitling - would evaluate how many subtitles start/end at the same time in hypothesis and reference, within a certain tolerance interval. That could also be interesting!

I get these numbers for some file I had lying around:

{
    "time_span_accuracy": 0.842,
    "time_span_precision": 0.858,
    "time_span_recall": 0.821,
    "time_span_f1": 0.839
}

If I take a completely different/wrong SRT file as reference I get:

{
    "time_span_accuracy": 0.511,
    "time_span_precision": 0.451,
    "time_span_recall": 0.485,
    "time_span_f1": 0.467
}

So very first impression is that those metrics could indeed be useful.