Model-based analysis of noisy musical recordings with application to audio restoration

Abstract

This thesis proposes digital signal processing algorithms for noise reduction and enhancement of audio signals. Approximately half of the work concerns signal modeling techniques for suppression of localized disturbances in audio signals, such as impulsive noise and low-frequency pulses. In this regard, novel algorithms and modifications to previous propositions are introduced with the aim of achieving a better balance between computational complexity and qualitative performance, in comparison with other schemes presented in the literature. The main contributions related to this set of articles are: an efficient algorithm for suppression of low-frequency pulses in audio signals; a scheme for impulsive noise detection that uses frequency-warped linear prediction; and two methods for reconstruction of audio signals within long gaps of missing samples.

The remaining part of the work discusses applications of sound source modeling (SSM) techniques to audio restoration. It comprises application examples, such as a method for bandwidth extension of guitar tones, and discusses the challenge of model calibration based on noisy recorded sources. Regarding this matter, a frequency-selective spectral analysis technique called frequency-zooming ARMA (FZ-ARMA) modeling is proposed as an effective way to estimate the frequency and decay time of resonance modes associated with the partials of a given tone, despite the presence of corrupting noise in the observable signal.