Complexity Reduction of WSOLA-Based Time-Scale Modification Using Signal Period Estimation

Abstract

In this paper, we propose a computational complexity reduction method for a waveform similarity overlap-and-add (WSOLA) based time-scale modification (TSM) algorithm using signal period estimation. In the proposed method, a signal period is estimated from the normalized cross-correlation. An optimal shift, a maximally similar point, of WSOLA for the current frame can be estimated from the estimated period obtained from the previous frame. Then, we reduce the search range for calculating the normalized cross-correlation around the estimated optimal shift instead of calculating for the full search range. In this manner, we can reduce the computational complexity required for normalized cross-correlations, which dominates most of the complexity in WSOLA. It is shown from experiments that the proposed method gives a relative complexity reduction of 56% for the WSOLA-based TSM algorithm while maintaining speech quality.