The timestamp of model 'interspeech21' is incorrect

[owaski]

I run the following command:

python -m allosaurus.run --timestamp=True -i sample.wav -m interspeech21

and it gives me

0.040 0.025 ɑ
0.080 0.025 l
0.100 0.025 ʌ
0.120 0.025 s
0.140 0.025 o
0.170 0.025 ɹ
0.180 0.025 ə
0.200 0.025 s

This is incorrect for the sample audio. Seems the window shift is set wrongly.

[SlistInc]

I am struggling with the timing as well. Is anybody aware of any library able to do a forced alignment of phonemes based on the input from allosaurus? I would really appreciate any input and tipps on how I can improve the output from allosaurus.

[journeytosilius]

I am also looking for something like this

[xinjli]

Hi guys, sorry I was a bit busy with other projects and my internship in the last few months and did not have time to look at it.

I forgot to count the subsampling factor from the conv layer, i fixed it in the latest commit.

[kzgajos]

A very useful library -- thank you for creating it.
I also have a timing issue. The onset of the phonemes seems to be reported correctly, but the duration of each shows as 0.045 regardless of how long each phoneme actually is. I need to detect pauses so accurate durations would be very helpful. Here's the output I get:

0.840 0.045 ʔ
0.870 0.045 a
0.900 0.045 l̪
0.960 0.045 t̪
0.990 0.045 ɒ
1.080 0.045 k͡p̚
1.140 0.045 a
1.260 0.045 t̪
1.320 0.045 ɒ
1.380 0.045 t̪
1.440 0.045 ɒ
1.470 0.045 k

[emorling]

i assumed it was because its returning the most likely phoneme at the 0.045 interval?