Obtained two promising WaveUnet models with following metrics:
8kHz: PESQ + 0.3(average), +0.3(0 SNR),
SDR +6.3(average), +9.5(0 SNR), 15.98 MMACs [Best Metrics]
8kHz + Dilations + DepthWise: PESQ + 0.3(average), +0.3(0 SNR),
SDR +6.2(average), +9.1(0 SNR), 7.49 MMACs [Smallest MAC count]
Obtained 16kHz model which is not smallest in terms of MACs, nor best in metrics, but perceptual quality is better due to higher sampling rate:
16kHz + Dilations + DepthWise: PESQ + 0.3(average), +0.3(0 SNR),
SDR +5.4(average), +8.2(0 SNR), 14.6 MMACs [Better perceptually]
Заказчик: Polyn Technology
Технологический стек: Python (PyTorch, scipy, librosa)