Performance Evaluation of CS Based Speech Enhancement using Adaptive and Sparse Dictionaries

Abstract

Traditional speech enhancement techniques proposed are limited to various types of stationary noise and lack robustness to some forms of real noise which are in general non stationary. Recently sparse dictionary based techniques have been applied to effectively handle the back ground noise in speech. Dictionary models in speech enhancement assume a specific type of a priori knowledge about the speech and noise signals. Two popular methods of dictionary models use commonly adapt an alternating optimization strategy in which the representation of a signal is fixed and the elements in the dictionary are learned, while the other find the sparse representation by keeping the dictionary elements fixed. Compressive sensing (CS) is an emerging technique that promises to effectively recover a sparse signal from very few random measurements than its dimension. A CS recovery algorithm, reconstructs a speech signal which is structured, while eliminating the unstructured noise. Encouraged by this emerging technique, the present work briefly reviews the performance of CS in speech enhancement using fixed dictionaries. In this paper, various sparse domain and sensing matrices and different combined transform domain (dictionary) pairs that satisfy incoherence condition has been investigated for their feasibility to perform speech enhancement. Extensive experiments are carried out to investigate the performance of the sparse transforms under different practical noise conditions taken from NOIZEUS database using various objective and subjective measures. The results obtained are very encouraging and helpful in the selection of dictionary suitable for CS based speech enhancement technique in practical application-based noise reduction.