Algorithm info: Using colab inference script (https://github.com/jarredou/FlashSR-Colab-Inference)
Mixture 036 and 047 were unusable during inference. With mixture 036, the cutoff is too low to be processed by FlashSR, a problem that is shared with AudioSR. As for mixture 047, not sure.
Because of this, the final score on all metrics is probably going to be lower than it could be.
First time I didn't think to resample every result back to 44100hz, did it by the 2nd attempt.
-Vinctekan
Metrics:
Metric sdr for restored: 16.5036
Metric si_sdr for restored: 16.3752
Metric l1_freq for restored: 31.2364
Metric log_wmse for restored: 14.7495
Metric aura_stft for restored: 8.6471
Metric aura_mrstft for restored: 21.8692
Metric bleedless for restored: 33.4487
Metric fullness for restored: 42.3405
Date added: 2025-02-01 |
