Yasas - Yet another sound analysis-synthesis method

This paper describes a novel analysis-synthesis method based on estimation and comparison between spectral methods that are specifically designed for modeling of random signals. These methods are an alternative to the discrete Fourier transform, providing high-resolution, smooth and easily interpretable spectral estimation over short signal frames. It is generally acknowledged that musical signals that exhibit both periodicity and variations can be modeled as sinusoidal components with various extents of modulation (generally characterized as band-limited frequency components), and remaining noise or wide-band frequency components that describe unvoiced and other noisy parts of the signal spectrum. Using single frame Fourier analysis makes is difficult to tell if spectral energy at a particular frequency is due to noise or sinusoidal components. Existing methods for such decomposition usually consider properties of Fourier phases, which are noisy. Our method decomposes a signal into sinusoidal, modulated sinusoidal and noise components based on a comparison between two spectral representations, namely autoregressive (AR) and minimum variance distortionless response (MVDR), both calculated from linear prediction coefficients (LPC) in a time varying manner. Using different optimal properties of each model, we develop estimators for frequencies and amplitudes (spectral envelope) of sinusoidal components (spectral lines) in noise and derive a "noisality" index that assigns different weights to contributions of sinusoidal and noise components at every frequency. Examples of synthetic and real sounds are presented in the paper.