Word timings - is there a way of knowing how long the gaps between words are?

[Mark Boas]

Hi All! 👋

I'm new to this forum, but have been following progress on Aeneas for a while now.

I think it's a great tool and this looks like a great community :)

We're thinking about using Aeneas in a number of projects, which is exciting.

We will need word level timings so I'm using 1.7.3 and I separated the text file I'm submitting so each word is on a separate line and am using the flag --presets-word to get more accurate timings.

If I'm understanding the docs correctly adding --presets-word should handle the MFCC nonspeech masking for me. Perhaps I'm misunderstanding the definition of a multi line text format or need to fine tune manually.

My text file looks like this:

I'm
used
to
the
idea
of
dying
while
I
have
no
desire
to
die
for
the
like's
of
you.

And I'm trying to align with this file: http://www.barbneal.com/wp-content/uploads/kirk07.mp3

Basically I run something like:

python -m aeneas.tools.execute_task \
    kirk07.mp3 \
    kirk07.txt \
    "task_language=eng|os_task_file_format=json|is_text_type=plain" \
    kirk07map.json --presets-word

The only issue is that I'm not seeing gaps between words in the outputted JSON timings. ie the end time of a word matches the start time of the next word, which is especially problematic when there are large pauses in the speech.

Here's an example of the output I get:

  {
   "begin": "1.240",
   "children": [],
   "end": "1.320",
   "id": "f000007",
   "language": "eng",
   "lines": [
    "dying"
   ]
  },
  {
   "begin": "1.320",
   "children": [],
   "end": "1.800",
   "id": "f000008",
   "language": "eng",
   "lines": [
    "while"
   ]
  },

There's a significant pause between "dying" and "while" which is not reflected in the output, and actually "while" occurs sometime after 4 seconds, although I wonder if that is a separate issue.

I've looked back through some of the discussions in this forum but to no avail, any suggestions are gratefully received.

Thanks in advance!

Mark