Matrix Study of Feature Compression Types and Instrumental Speech Quality Metrics in Ultra-Light DNN-Based Spectral Speech Enhancement

Spectral speech enhancement with deep neural networks (DNNs) often relies on linear, logarithmic, or more rarely power-law compressed spectral features. Our study therefore investigates the effects of these compression functions in terms of instrumental performance metrics in an otherwise comparable DNN framework built upon the convolutional denoising autoencoder. Using the established data from the deep noise suppression (DNS) challenge in all configurations, this paper experimentally relates the available compression functions to the typical instrumental performance metrics such as DNS-MOS, wideband PESQ, STOI, and segmental SNR. In this matrix of system options and speech quality metrics, cleared from other influences, it turns out that power-law compression combines the different amenities of linear and logarithmic compression and even overcomes a former dual-compression system with linear and logarithmic subsystems combined through a meta-network.