We present here an Harmonic+Noise Model (HNM) for speech synthesis. The noise part is represented by an autoregressive model whose output is pitch-synchronously modulated in energy. The harmonic part of the signal is represented by a sinusoidal model. This paper compares different methods for separating these two components. We then propose a method for the estimation of the sinusoidal parameters derived from the ABS model [8] and evaluate different models for the analysis/synthesis of the stochastic part proposed in the literature.