Task can be found here.

Based on start code from here.

Should be used in pair with Praat.

I created 2703 text segmentation files from LibriSpeech dev-clean dataset using provided LSTM acoustic model and watched some of them in Praat special software. Examples can be found in results_samples folder.

Warning: there were several (less then 100) errors of unknown origin while creating segmentations.

We can see analisation example for sample 174-50561-0:

Here we can see that there is a small (0.22 seconds) delay between first sound peak and first segmented sound. Also the last one is 0.15 seconds after last significant sound. It can be related to convolution downsampling and time needed for the model to "realize" heard piece of sound. This effect is stable across many seen samples.

Alignments are pretty accurate, but there were several mislabelings in most of the predictions. But it is quite ok. Also, on this scale in most of the applications, sounds alignment and duration estimation quality is excellent.