Abstract

Speech enhancement is a widely adopted technique that removes the interferences in a corrupted speech to improve the speech quality and intelligibility. Speech enhancement methods can be implemented in either time domain or time-frequency (T-F) domain. Among various proposed methods, the time-frequency domain methods, which synthesize the enhanced speech with the estimated magnitude spectrogram and the noisy phase spectrogram, gain the most popularity in the past few decades. However, this kind of techniques tend to ignore the importance of phase processing. To overcome this problem, the thesis aims to jointly enhance the magnitude and phase spectra by means of the most recent deep neural networks (DNNs). More specifically, three major contributions are presented in this thesis.

First, we present new schemes based on the basic Kalman filter (KF) to remove the background noise in the noisy speech in time domain, where the KF acts as joint estimator for both the magnitude and phase spectra of speech. A DNN-augmented basic KF is first proposed, where DNN is applied for estimating key parameters in the KF, namely the linear prediction coefficients (LPCs). By training the DNN with a large database and making use of the powerful learning ability of DNN, the proposed algorithm is able to estimate LPCs from noisy speech more accurately and robustly, leading to an improved performance as compared to traditional KF based approaches in speech enhancement. We further present a high-frequency (HF) component restoration algorithm to extenuate the degradation in the HF regions of the Kalman-filtered speech, in which the DNN-based bandwidth extension is applied to estimate the magnitude of HF component from the low-frequency (LF) counterpart. By incorporating the restoration algorithm, the enhanced speech suffers less distortion in the HF component. Moreover, we propose a hybrid speech enhancement system that exploits DNN for speech reconstruction and Kalman filtering for further denoising. Two separate networks are adopted in the estimation of magnitude spectrogram and LPCs of the clean speech, respectively. The estimated clean magnitude spectrogram is combined with the phase of the noisy speech to reconstruct the estimated clean speech. A KF with the estimated parameters is then utilized to remove the residual noise in the reconstructed speech. The proposed hybrid system takes advantages of both the DNN-based reconstruction and traditional Kalman filtering, and can work reliably in either matched or unmatched acoustic environments.

Next, we incorporate the DNN-based parameter estimation scheme in two advanced KFs: subband KF and colored-noise KF. The DNN-augmented subband KF method decomposes the noisy speech into several subbands, and performs Kalman filtering to each subband speech, where the parameters of the KF are estimated by the trained DNN. The final enhanced speech is then obtained by synthesizing the enhanced subband speeches. In the DNN-augmented colored-noise KF system, both clean speech and noise are modelled as autoregressive (AR) processes, whose parameters comprise the LPCs and the driving noise variances. The LPCs are obtained through training a multi-objective DNN, while the driving noise variances are obtained by solving an optimization problem aiming to minimize the difference between the modelled and observed AR spectra of the noisy speech. The colored-noise Kalman filter with DNN-estimated parameters is then applied
to the noisy speech for denoising. A post-subtraction technique is adopted to further remove the residual noise in the Kalman-filtered speech. Extensive computer simulations show that the two proposed advanced KF systems achieve significant performance gains when compared to conventional Kalman filter based algorithms as well as recent DNN-based methods under both seen and unseen noise conditions.

Finally, we focus on the T-F domain speech enhancement with masking technique, which aims to retain the speech dominant components and suppress the noise dominant parts of the noisy speech. We first derive a new type of mask, namely constrained ratio mask (CRM), to better control the trade-off between speech distortion and residual noise in the enhanced speech. The CRM is estimated with a trained DNN based on the input noisy feature set and is applied to the noisy magnitude spectrogram for denoising. We further extend the CRM to the complex spectrogram estimation, where the enhanced magnitude spectrogram is obtained with the CRM, while the estimated phase spectrogram is reconstructed with the noisy phase spectrogram and the phase derivatives. Performance evaluation reveals our proposed CRM outperforms several traditional masks in terms of objective metrics. Moreover, the enhanced speech resulting from the CRM based complex spectrogram estimation has a better speech quality than that obtained without using phase reconstruction.