Classification and separation techniques based on fundamental frequency for speech enhancement

Abstract

This thesis is focused on the development of new classification and speech enhancement algorithms based, explicitly or implicitly, on the fundamental frequency (F0). The F0 of speech has a number of properties that enable speech discrimination from the remaining signals in the acoustic scene, either by defining F0-based signal features (for classification) or F0-based signal models (for separation). Three main contributions are included in this work: 1) an acoustic environment classification algorithm for hearing aids based on F0 to classify the input signal into speech and nonspeech classes; 2) a frame-by-frame basis voiced speech detection algorithm based on the aperiodicity measure, able to work under non-stationary noise and applicable to speech enhancement; 3) a speech denoising algorithm based on a regularized NMF decomposition, in which the background noise is described in a generic way with mathematical constraints.