search

auspicious3000 / contentvec

speech self-supervised representations
MIT License
434 stars 32 forks source link

Question about released model and speaker information #15

Closed asr-pub closed 11 months ago

asr-pub commented 11 months ago

Hello, Thank for your great work.

  1. The released model were trained on LibriSpeech 960hrs ?
  2. In the paper, the teachers are discrete semantic tokens, How to measure the Speaker info. in teachers ? image
auspicious3000 commented 11 months ago

This is only a conceptual curve of how speaker information changes through the network layers. You can roughly "measure" the speaker information by, for example, measuring the speaker classification accuracy.

asr-pub commented 11 months ago

This is only a conceptual curve of how speaker information changes through the network layers. You can roughly "measure" the speaker information by, for example, measuring the speaker classification accuracy.

OK, Thank u