Techniques For Enhancement Of Noisy Speech In Adverse Environments

The enhancement of speech signal corrupted by additive noise has attracted many re-searchers in recent years. They are two major parts involved in the process of designing speech enhancement systems:the estimation of the clean speech spectrum and the estimation of the noise spectrum. This thesis addresses both issues.The first contribution of this thesis is the development of an improved speech enhancement technique based on a modified Minimum Mean-Square Error Log-Spectral Amplitude (MMSE-LSA) and on an efficacious estimator for the a priori Speech Absence Probability (SAP) which is directly applied to the modified MMSE-LSA as it requires such an estimate. Several popular techniques based on LSA estimator use either a small fixed value for the a pri-ori SAP or small fixed value for its parameters. Since the use of static values for the a priori SAP has shown not to be optimal and unsuitable as the computation of the a priori SAP is in-fluenced by noise-only regions, it is preferable to use dynamic parameters to compute the a pri-ori SAP estimator so it can continuously be updated in each frame. In this thesis, we propose two dynamic factors for the computation of the a priori SAP:a smoothing-update factor and a factor connected with the k-th spectral component. The smoothing-update factor which is based on a decision made in frequency band whether speech is present or absent, is computed by re-cursively averaging past spectral values of the a priori SAP. While the factor related to the k-th spectral component relies on a priori signal-to-noise ratio (SNR). The decision on each fre-quency band whether speech is absent or present is determined by computing and comparing the conditional probabilities of the noisy speech spectrum. Two objective measures in addition to study of speech spectrograms were selected to evaluate the proposed enhancement technique. Evaluation results attest the superiority of this method regarding the quality of speech, the re-duction of background noise and the amount of distorted speech introduced.The second contribution of this thesis is the development of a simple alternative to existing noise estimation methods for adverse environments. The algorithm is based on an improved method for tracking the minimum of noisy speech by continuously averaging past spectral val-ues of the noisy speech and on a dynamic smoothing update factor computed to control the up-date of the estimated noise spectrum during speech presence. The dynamic smoothing update factor is computed based on a decision made in frequency band whether speech is present or absent. The distinction between speech present and speech absent is determined by computing the difference of the noisy speech power spectrum to its local minimum. Unlike other methods computed under speech presence uncertainty, our algorithm does not explicitly use a speech-presence probability. The evaluation of the proposed approach using NOIZEUS data-base shows that, when integrated into speech enhancement system, it can achieve better speech quality and significant noise reduction while minimizing the amount of speech distortion and musical noise in the processed speech signal.