HNM: a simple, efficient harmonic+noise model for speech

Abstract

HNM, a new analysis/modification/synthesis model based on a harmonic+noise representation of the speech signal is presented. The HNM model has several specificities: (1) HNM assumes the speech signal to be composed of a deterministic part and of a stochastic part, (2) the deterministic part is assumed to contain only harmonically related sinusoids with linearly varying complex amplitudes, and (3) the stochastic part is modeled in both the time domain and the frequency domain. The pitch-synchronous analysis technique makes use of a coarse estimate of the pitch and simultaneously calculates the various parameters of the model and refines the pitch estimate. Because the signal is decomposed into a deterministic and a stochastic part, different modification methods can be applied to each part, yielding more natural resyntheses.<>