Confidence in alignment, robustness to incorrect transcript

[gschalk]

Hi,

Aeneas absolutely rocks!!

I have two questions to the group today:

1) Is a way for me to learn the confidence in the alignment of a particular section of text.  In other words, if I want to eliminate those sections that are probably incorrectly aligned, is there a way for me to kick them out based on some metric of performance/fit/etc. produced by aeneas?  I understand that, at the most basic level, one could study the fit or offset from diagonal path in the DTW, but I am not sure if any of these variables or other performance metrics may be available on the command line.

2) I have long audio files whose transcripts occasionally miss maybe 15 seconds of spoken text.  Sometimes aeneas is robust to the those transcription problems, sometimes it's not.  I understand I can make the DTW more "loose," and there are other parameters I can play with, but is there some principled way of thinking about what kind of parameters I should be changing in what ways to get optimal results (kind of like the automated head/tail detection where I provide a max head/tail time)?

Thanks for any help you may provide.

Gerwin