Particle filtering methods for the enhancement of speech corrupted by additive noise

Abstract

In this work, we study the application of particle filtering (PF) algorithms to the problem of speech enhancement. The goal of the thesis is to devise PF algorithms that will enhance speech signals corrupted by additive noise, and to evaluate their performance via comparisons with other existing algorithms based on several quality measures. Speech enhancement, or noise reduction, is an important problem in many applications, such as telephony and telecommunications in general, sound recording, human-coaching interface (where speech recognition is important), etc. Even though many algorithms already exist for speech enhancement, there is still very much work to do, especially in terms of intelligibility. In many cases, it may be easier to understand the original, noisy speech rather than the processed, "cleaned-out" one. In other cases, the residual noise may be too annoying to carry out a comfortable conversation. In this context, new approaches for the denoising of speech are welcome. As a first contribution, a practical approach to deriving simple Rao-Blackwellised Particle Filters (RBPFs), which was developed in parallel with a theoretic review of PFs, is presented. In addition, a novel algorithm, called the modified Rao-Blackwellised Particle Filter (RBPF), is proposed to reduce the computational load of regular RBPFs. Several new speech enhancement methods using particle filters are also derived, and shown to outperform some other existing PF-based algorithms. Accessorily, a novel strategy to extend their range of application to colored noise is explained and applied. Comparatively to the other types of enhancement algorithms tested (including spectral subtraction, signal subspace, dual extended Kalman filter, perceptually constrained Kalman filter, dual perceptually constrained unscented Kalman filter) we find that the particle-filter based algorithms presented have the advantage of not introducing any musical noise. Furthermore, in the conditions of our experiments, using several objective measures we find that they are able to compete with and outperform most of the other algorithms tested. Using these measures and based on informal listening, we highlight their advantages---naturalness of the enhanced speech, low intrusiveness of the non-musical residual noise, very good performance at high SNR, flexibility---and their main limitations---intraspeech residual noise "modulated" by the speech, computational burden. Considering how flexible and parametrizable PFs are, there is a strong potential for further improvement.