Inquiry about Reproducing libri2mix_eend_ss Model - Performance Not Meeting Expectations (8.54%)
Dear Author,
I hope this message finds you well! I'm a developer/researcher who is currently trying to reproduce your model libri2mix_eend_ss shared on Hugging Face. I've read the provided usage instructions (the running script based on ESPnet2) and really appreciate your valuable open-source work!
During the reproduction of your libri2mix_eend_ss model, I have strictly followed the steps in the documentation to configure the environment and execute the script. However, the model performance has consistently only reached 8.54%, which is lower than expected. If you're available, could you kindly share some key tips related to reproduction? For example, the specific dependency versions to pay attention to, supplementary instructions for dataset downloading, the detailed structure of the training configuration file, and any pitfalls you summarized during training/testing? My email address is: leijiegong1@gmail.com. Looking forward to your valuable guidance!
Regardless of whether I receive a response, I'm extremely grateful for your willingness to open-source the model and share technical details—it’s a great help for our learning and research! Looking forward to your valuable suggestions.
Best regards,
Gong Leijie