brentspell / hifi-gan-bwe

Unofficial implementation of HiFi-GAN+ from the paper "Bandwidth Extension is All You Need" by Su, et al.
MIT License
206 stars 26 forks source link

Performance Evalution Metric #6

Closed saivinaypsv closed 2 years ago

saivinaypsv commented 2 years ago

Hi brentspell,

What is the PERFECT/BEST evaluation metric for BWE ? POLQA can be used for BWE evalution metric or not?

brentspell commented 2 years ago

If you take a look at the original paper, the authors use Peak Signal-to-Noise Ratio (PSNR) and Log-Spectral Distance (LSD) as objective evaluation metrics. However, they note that these aren't great indicators of perceptual quality, so they collect Mean Opinion Scores (MOS) using Amazon Mechanical Turk. I haven't done any of these evaluations myself.

I have had some success in the past with the PESQ evaluation metric for vocoders. However, using it requires downsampling to 16kHz, which probably won't work well for bandwidth extension.