Using multi-scale product spectrum for single and multi-pitch estimation

Abstract

The authors present an algorithm for pitch estimation including voiced/unvoiced decision in the case of a noisy speech and when two speakers are talking simultaneously. The approach is based on the spectral multi-scale product (SMP) analysis of the sound mixture. SMP is the spectrum of the product of three successive wavelet transform coefficients of the speech. The wavelet used for SMP analysis is the quadratic spline function. The proposed method is compared with other state-of-the-art algorithms. It is robust in the presence of a noise and permits the pitch estimation of the dominant speech and the concurrent one from the sound mixture with high accuracy.