INTRODUCTION

In the field of audio signal processing, sinusoidal modeling is a fundamentally important signal representation. It finds applications in audio coding, speech enhancement, sound source separation, and pitch tracking. All these applications benefit from high quality sinusoidal parameters. In conventional sinusoidal modeling [1,2], these parameters are defined as frequency, amplitude and phase of each frequency component. Other efforts have focused on non-sinusoidal parameters such as bandlimited noise [3] or transients [4]. Efforts have also been made to expand or improve sinusoidal parameters. Matching-pursuit [5] seeks to identify exponentially decaying envelopes of quasistationary sinusoids. Linear frequency chirp rates of frequency components have been estimated [6], when constant- or Gaussian-amplitude time envelopes are present. A dynamic vocoder model [7] has been used to estimate very accurate sinusoidal trajectories within each frame, but at great computational expense. We present a new technique that efficiently and robustly obtains the linear chirp parameter. The current estimation technique is robust to phase shift, and to time domain amplitude modulation.