Open bigTexas opened 5 years ago
francoisgermain
commented
5 years ago Hi. Thanks for your interest, and sorry to hear about your issue. I am pretty busy at the moment and I no longer have access to a GPU machine. How urgently do you need an answer? In case that helps in the meantime, I believe I have somewhere the audio files that were generated (and scored) for the paper results. Let me know.
bigTexas
commented
5 years ago Thanks for your replying @francoisgermain ! It will be very helpful if you can point me to the already generated enhanced audios for validation dataset. When you get time later, I appreciate if you can run the inference code and recompute the metrics for the generated audio. For your information, the number I got is computed on 1080Ti GPU and none of your code is changed.
aleXiehta
commented
3 years ago Hi, I have also met this problem when running the inference code using pre-trained weights provided in the repo and generated the enhanced utterances.
The evaluation tool I'm currently using is from here, and the scores I've got are: | PESQ | CSIG | CBAK | COVL | SSNR Deep Feature Loss (paper) | - | 3.86 | 3.33 | 3.22 | - Deep Feature Loss (code) | 2.58 | 3.80 | 2.72 | 3.19 | 0.60
It seems that the main differences are CBAK and SSNR. I will be very grateful if you can give me any support for this. Thank you!
First thanks for providing the code. I tried to test the algorithm by running the inference (python senet_infer.py) on the validation dataset. The metrics I obtained is avg_sig=3.785319, avg_bak=3.273069, avg_ovl=3.141790 while the number reported in the paper is: avg_sig=3.86, avg_bak=3.33, avg_ovl=3.22. I'm wondering whether I missed something. I didn't change any code and use exactly the same evaluation code mentioned in the paper. FYI, when I tested on the noisy data (no enhancement is applied), the numbers I got is the same as that in paper. Any idea?