Optimization of Feature and Loss Exponents for Lightweight DNN-based Binaural Speech Enhancement

Abstract:
Lightweight DNN-based binaural speech enhancement (BSE) requires careful system design to solve the trade-off between resource efficiency and performance. Using eight instrumental metrics for noise reduction, speech intelligibility and audio quality, the BSE components such as feature compression and loss function are optimized in a joint exploration of power-law and mean p-power error exponents, providing guidance on BSE design used for implementation and comparison of systems. The study uses standard 16 kHz and advanced 24 kHz sampling rates.