Perceptually Motivated Approaches to Audio Signal Enhancement: Broadband Noise Reduction via Bayesian Modeling of Time-frequency Coefficients

• Activity Number: 147
• Type: Topic Contributed
• Date/Time: Monday, August 9, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Patrick J. Wolfe*+
• Companies: University of Cambridge
• Keywords: regression ; model selection ; Bayesian estimation ; regularization ; risk theory ; Markov chain Monte Carlo

Abstract

This paper describes perceptually motivated statistical models for audio data, formulated to effect broadband noise reduction in natural sound signals such as speech and music. In contrast to most other Bayesian approaches to date, however, the methodology pursued here involves modeling of time-frequency coefficients--thus forming a natural yet novel extension of the techniques currently employed in many audio signal processing applications. Standard methods are interpreted from a Bayesian viewpoint and consequently extended to develop new, computationally efficient algorithms for online noise reduction. Working within this context, Bayesian risk theory is then employed in conjunction with perceptual optimality criteria to devise noise suppression rules motivated by psychoacoustic considerations. Masked thresholds in turn provide a basis for perceptual cost functions, under which minimum-risk spectral amplitude estimators are derived. Lastly, hierarchical models employing MCMC methods are developed to enhance speech signals degraded by noise, in which case meaningful prior information is shown to aid in the solution of the resultant ill-posed inverse problem via regularization.