Speech Periodicity Enhancement Based on Transform-domain Signal Decomposition and Robust Pitch Estimation

Periodicity is an important attribute of speech signals. It is an essential element of tonal languages, where the meaning of a word is determined by the pitch contour. Speech periodicity enhancement is the process of restoring waveform periodicity of noise-corrupted speech, in order to improve human perception of pitch and tone in noisy environments.;This thesis presents a novel approach to speech periodicity enhancement. The enhancement is achieved through periodic-aperiodic decomposition of the linear prediction residual signal in a transform domain. Transform coefficients that represent the periodic component are amplified to enhance the periodicity, and those coefficients representing the aperiodic components are attenuated to suppress the noise. We propose and evaluate different methods of assigning coefficient weights for periodicity enhancement. These methods include simple fixed weights, adaptive weights, and transform-domain Wiener filtering.;As a key component for periodic-aperiodic decomposition, a novel method of robust pitch estimation is developed. The temporally accumulated peak spectrum is proposed as a robust representation of speech harmonics. Gaussian mixture model is employed to model the effect of noise on the peak spectrum. Pitch estimation is formulated as a problem of l1-regularized maximum likelihood estimation, in which prior information is exploited. Two convex optimization approaches are developed to solve the associated non-convex optimization problem. The proposed pitch estimation method significantly outperforms the conventional methods. It attains high estimation accuracy for various types of noise at very low signal-to-noise ratio (e.g., -5 dB).;Experimental results confirm that with the proposed approach of periodicity enhancement, speech harmonic structure and waveform periodicity can be effectively restored. Compared with other speech and periodicity enhancement methods evaluated in this study, the proposed method can produce speech outputs with noticeably higher quality in terms of different objective measurements, such as SNR and PESQ.