search

microsoft / DNS-Challenge

This repo contains the scripts, models, and required files for the Deep Noise Suppression (DNS) Challenge.
Creative Commons Attribution 4.0 International
1.11k stars 412 forks source link

Question about the DNSMOS on clean speech #189

Open SaoYear opened 2 months ago

SaoYear commented 2 months ago

Hi,

I have a question about the updated version of DNSMOS.835. I testified the clean speech performances on the DNSMOS (non-personalized), but found the model would give a score only about 3.2-3.3 for even clean speech.

Implementation: I directly used the dnsmos_local.py and the .onxx files provided in this repo.

Here are some results (all clean speech):

  1. LibriTTS (~25k wav files): 3.23
  2. DNS-challengeI test set, synthetic clips without reverb: 3.28
  3. CHIME4 simulated test set: 3.31

It seems that the upper bound of the current version of DNSMOS is about 3.3. My question is that, is this the intention of the MOS score?

Many thanks in advance!

hollandloprabbit commented 1 month ago

I have the same problem, any explanation?