How to get alignment metrics from result?

[bonsai-byte]

When model.align is invoked, it sometimes prints warning messages for how many segments failed to get aligned, or how many words could not be aligned at the end of an audio clip. I am aligning song audio with lyrics, and certain genres do not benefit from vad=True. I want to run with vad=True and vad=False; based on the alignment metrics, I'll pick the one with a higher likelihood. How can I do that?

[jianfch]

The fail count is not stored in the result but you can recount it. The segment/word is considered as failed if the end-start<=0.
Note that this different from the duration attribute of the segment/word because duration is end-start rounded to 3 decimals.

stable-ts/stable_whisper/alignment.py

Line 702 in 6826786

if fail_segs := len([None for s in result.segments if s.end-s.start <= 0]):

The words that failed to aligned at the end of an audio clip have their end and start set to the audio duration.

# failed segments / total segments
failed_segments = len([None for seg in result.segments if seg.end-seg.start <= 0])
print(f'Failed segments: {failed_segments}/{len(result.segments)}')

# failed words / total words
all_words = result.all_words()
failed_words = len([None for word in all_words if word.end-word.start <= 0])
print(f'Failed words: {failed_words}/{len(all_words)}')

# remaining words that failed to align / total words
last_failed_words = 0
for word in reversed(all_words):
    if word.end == word.start:
        last_failed_words += 1
    else:
        break
print(f'Failed remaining words: {last_failed_words}/{len(all_words)}')