On Sparsity of Speech Features with Ladder Autoencoders for Multi-Speaker Separation

Abstract

The multi-speaker separation mechanism consists of speech feature extraction and temporal coherence. In this study, a speech feature extraction is developed, and the reconstructed-speech quality is evaluated with different degrees of sparsity. Speech feature extraction is implemented on ladder autoencoders with branches embodying a sparse encoder-decoder model where the autoencoders are trained with the WSJ0-2mix English Corpus. An evaluation indicates the stability of the reconstructed-speech quality, with a signal-to-distortion ratio of >5 dB in the sparseness range of 0.4–0.7. The results suggest the applicability of the feature extraction method to the investigation of temporal coherence.