A temporal matching method for pitch determination from noisy speech signals

Abstract

A new method based on temporal matching is presented in this paper for pitch determination from noisy speech signals. At first, a variable-length average magnitude sum function (VLAMSF) that exhibits a periodicity similar to the pitch period of the voiced speech, has been proposed. We argue that the discrete cosine transform (DCT) power spectrum of the VLAMSF is capable of revealing an estimate of a pitch-harmonic (PH) more accurately even in a heavy noisy scenario. Then, exploiting the extracted PH, we formulate an impulse train with a variable-period that is used to temporally match the periodicity of the proposed VLAMSF for pitch determination. It has been shown through extensive simulations using the Keele database that our new approach consistently outperforms the other existing methods especially at low signal-to-noise ratios (SNRs).