Study Of Speech Enhancement Based On Noise Estimation And A Priori SNR Estimation

Numerous speech processing devices (e.g. mobile communication systems and digital hearing aids systems) which are often used in environments with high levels of ambient noise such as public places and cars in our daily life. Speech enhancement methods can be used to increase the quality of these speech processing devices. Correct noise power spectrum estimation and a priori signal-to-noise ratio (SNR) estimation are all essential to good quality of the enhanced speech. This paper will focuses on noise estimation and apriori SNR estimation based on the assumption of a slowly varying noise environment and the noisy speech alone is available.The main works of the paper can be summarized as follows:At first, a new noise power spectrum estimation algorithm which is namely weighted recursive averaging is presented for non-stationary noise environments. The new proposed algorithm continuously updates the estimated noise by weighted noisy speech with a time-varying frequency-dependent smoothing factor. To avoid noise overestimates in the weak speech regions following strong speech, a projecting smoothing method which makes more accurate SNR estimate available for smoothing factor and weighting factor calculations is proposed. The proposed noise estimation algorithms do not wait for specific window time to update the noise estimate. Hence the update for varying noise power levels is much faster compared to most other algorithms and at the same time noise power spectrum is not overestimated. Objective experimental results indicate that the proposed noise estimation algorithms when integrated in speech enhancement yields a significantly improvement in the segmental SNR, lower segmental estimation error, and better Perceptual Evaluation of Speech Quality (PESQ) scores.Secondly, an improved projecting smoothing method by taking into account the correlation of inter-frequency bins in the observed signal is proposed, resulting in more accurately distinguishing noise and weak speech.When the new presented projecting smoothing method is combined into the weighted recursive averaging algorithm, the improved weighted recursive averaging algorithm is getted. The tracking delay and the overestimates are all considerably reduced compared to previous noise tracking algorithms. Objective experimental results and a subjective comparison show that the improved noise estimation algorithm when integrated in speech enhancement is preferred over other noise estimation algorithms. Finally, a novel approach which improves the a priori SNR estimation for speech enhancement in noisy environments is proposed. The new approach combines decision-directed estimation and predicted estimation, associated with the speech presence uncertainty. The employed predicted estimation incorporates the soft-decision scheme. In contrast to the conventional decision-directed estimation scheme, a further reduction of residual noise is achieved. The paper also shows that a special case of the proposed method degenerates to the decision-directed estimation with a weighting factor. The use of the proposed a priori SNR estimator leads to get higher segmental SNR improvement and PESQ scores.