Algorithm info: 1. BS Roformer (vocals, instrumental) ver 2024.04 (SDR vocals: 11.24, SDR instrum: 17.55)
2. BS Roformer (vocals, instrumental) ver 2024.08 (SDR vocals: 11.31, SDR instrum: 17.62)
3. MelBand Roformer (vocals, instrumental) Bas Curtiz edition (SDR vocals: 11.18, SDR instrum: 17.49)
4. Weights https://github.com/KimberleyJensen/Mel-Band-Roformer-Vocal-Model
stereo sequences --111111
[0, 1, 2, 4, 7, 9, 13, 23]
Estimations: gross distinctions
N25e10
Instrum N25e-10
halfwave 2048, 250, 332
My options were limited a little by Google Colab. It was running out of memory and crashing.
Hunter Hogan
Metrics:
Metric sdr for instrum: 17.6572
Metric si_sdr for instrum: 17.5638
Metric l1_freq for instrum: 40.3569
Metric log_wmse for instrum: 14.4029
Metric aura_stft for instrum: 13.1926
Metric aura_mrstft for instrum: 17.0620
Metric sdr for vocals: 11.4766
Metric si_sdr for vocals: 11.0816
Metric l1_freq for vocals: 39.8353
Metric log_wmse for vocals: 14.5148
Metric aura_stft for vocals: 8.9644
Metric aura_mrstft for vocals: 11.2615
Date added: 2024-10-22 |
