About Speech Enhancement scores on SUPERB leaderboard

Speech-Lab-IITM / CCC-wav2vec-2.0

Code for the method proposed in the paper:- ccc-wav2vec 2.0: Clustering aided Cross-Contrastive learning of Self-Supervised speech representations

MIT License

13 stars 3 forks source link

About Speech Enhancement scores on SUPERB leaderboard #1

Open leo19941227 opened 1 year ago

leo19941227 commented 1 year ago

Hi!

Thanks for the great repo! I find that CCC-wav2vec 2.0 performs especially well on SUPERB SE task, surpassing WavLM Large with a large margin. I am trying to reproduce it but I not yet successfully get the similar score. (PESQ around 3) May I ask how did you fine-tune CCC-wav2vec 2.0 on SUPERB SE? Specifically, what are the tuning hyper-parameters? (e.g. batch size and learning rate) Thanks!

leo19941227 commented 1 year ago

P.s. I tried the default batch size and learning rate 1e-4 and 5e-5, and adding --upstream_feature_normalize to apply layer norm on the representation, the best result I get is around 2.6 PESQ.