Seems not work on speechocean dataset

[louislau1129]

Hi, @felixkreuk , first thank you for open-sourced such good repo on unsupervised phoneme segmentation. Recently, I conduct several experiments on SpeechOcean 762 dataset, which is a standard speech scoring dataset.

1. First, I directly apply the provided pretrained boundary detection model on this corpus, and only found about 50% F1 and R value.
2. I suspect this may relates to the domain mismatch problem, so I try to re-train this boundary detection model on the SpeechOcean corpus from scratch, but still attains about 50% F1 and R value, it is far lag from the referenced force aligned boundary result.

The following result screenshot is about using the random initialized model (without start training) to predict:

The following result screenshot is about using the trained model to predict:

Any idea on fixing this issue? Thanks in advance!

[louislau1129]

(Ps: I have re-produce the reported result on timit using the code