Training bottleneck hparams

auspicious3000 / SpeechSplit

Unsupervised Speech Decomposition Via Triple Information Bottleneck

http://arxiv.org/abs/2004.11284

MIT License

636 stars 92 forks source link

Closed Merlin-721 closed 3 years ago

Merlin-721 commented 3 years ago

Can I get a confirmation on tuning the bottlenecks inline with Appendix B.4?

"The first operation is to increase the channel dimension of the encoder output" Does this refer to dim_enc, dim_enc_2 and dim_enc_3?
"The second operation is to increase the sampling rate of the down-sampled code" Does this refer to freq, freq_2 and freq_3?

auspicious3000 commented 3 years ago

Yes