An overlap-add free musical noise analysis-synthesis system

Abstract

An analysis-synthesis system for audio signals using sinusoidal model-based algorithms for harmonic component reconstruction and noise modeling for stochastic component synthesis has recently demonstrated the generation of high-quality synthetic signals. However current approaches, which mainly use phase randomization noise-driven source filter models with overlap-add techniques, have degraded the performance for signals consisting of highly time-localized events such as transients. First, the undesirable effects of window overlap-add analysis-synthesis on non-stationary signals are addressed, second, an overlap-add-free analysis-synthesis system is proposed based on a coherent loss compensation concept, and third, a noise synthesis model based on the perceptual mechanism is described using sinusoidal components randomly spaced in the frequency domain and novel phase prediction and phase correction techniques. The listening results illustrate our overlap-add free noise synthesis model has better performance than the conventional window overlap-add noise-driven filter models with phase randomization for highly time-localized signals.