Harmonic-plus-noise decomposition and its application in voiced/unvoiced classification

In this paper, we present an improved algorithm to decompose the harmonic and the noise components of voiced speech. The improvements make the method more accurate and robust by employing a harmonic extrapolation and a noise extrapolation in alternating iterative steps and by including a new pitch detection algorithm. This new technique has been found to improve both the convergence and accuracy of separation of the harmonic and the noise components. In separating the noise and the harmonic components, this improved harmonic-plus-noise (H+N) decomposition method provides many useful ways to measure the strength of voicing. Two such measures are investigated with respect to their ability to discern voiced and unvoiced segments of speech. They are the harmonic-to-noise energy ratio and the sub-band harmonic-to-noise energy ratio. Tests show that these measures perform more reliably and more robustly in comparison to classical measures such as the zero-crossing rate, the LPC prediction gain, the 1^st LP coefficient and the RMS energy