A novel speech reconstruction algorithm for DSR back-end

In this paper, a novel speech reconstruction algorithm for DSR back-end is proposed. The algorithm is based on the classic least-squares estimate, inverse short-time Fourier transform magnitude (LSE-ISTFTM) algorithm. Unlike the classic LSE-ISTFTM algorithm, initializing speech waveform with white noise, the proposed method reconstructs voiced and unvoiced speech waveform separately, initializing with a specific signal. Furthermore, the magnitude spectrum is inversed from MFCC with Moore-Penrose pseudo-inverse by Mel-scale weighting functions. The algorithm evaluation results show that the proposed Extended LSE-ISTFTM algorithm converges faster and more stable than the classic algorithm. The speech reconstruction results demonstrate that PESQ score of the proposed algorithm is higher than the classic LSE-ISTFTM algorithm and the DSR back-end method.