support for alignment output in tsv format

[contentnation]

Support of alignment data output.
Kind of matching on issue #364
Can be used as a base for #391 and #361
Runs text to speech 2 times, one for normal audio generation,
a second time for each word.
Since both produce different outputs and times, a correction is applied.
Not perfect, but "good enough". Both will self sync after each sentence, so only slight offset are created.

[vytskalt]

I've been trying this out. Looks like when using a long text some of the last words are being skipped in the alignment file.

[contentnation]

@vytskalt can you provide an example so I can debug/fix it?

[vytskalt]

@vytskalt can you provide an example so I can debug/fix it?

Yes, this is the command I'm running:

cat text.txt | piper --sentence-silence 0.5 -m en_US-ryan-high --output_file out.wav --alignment-data alignment.tsv

This is the text (random Reddit post): text.txt

In the alignment.tsv, 2 of the last words are missing.

[contentnation]

ok, it's not the length that is the issue, it's the content. For example: "musical/sport" will be spoken as 3 words. "in the" is mangled into one spoken word. My word/phoneme sync trips over this. Needs to be fixed, I have to find another way to sync.

[charlyhayoz]

Hi,

i pulled this pull request and make a build but the --ali gnment-data is not disponible in the executable "piper" in the install folder.

Am i missing something to make it work ?

Thanks (:

[contentnation]

It is only built into the python script, not in the c++ executable.

[charlyhayoz]

Make sense ! Thanks (:

[WilleIshere]

It is only built into the python script, not in the c++ executable.

Hi would it be possible to add it to the c++ exe? i am using windows which does not have the python version so i need it compiled. Dont know how to translate python to c++

[synesthesiam]

The w_ceil tensor from the original VITS models contains the phoneme timing information (when multiplied by the hop length of 256). Just routing this tensor through the model output is enough to get alignment information. Unfortunately, this requires re-exporting all of the existing voices unless we can find a way to get access to this tensor within the ONNX models.