Joint tracking of clean speech and noise using HMMs and particle filters for robust speech recognition

Abstract

We propose a dynamic joint tracking framework to monitor the clean speech signal and noise simultaneously in order to compensate the noisy features. The clean speech signal is tracked using an integrated algorithm based on both particle filters and hidden Markov models. The information available from speech tracking is used for tracking and estimating the noise parameters. The availability of dynamic noise information enhances the robustness of the algorithm in case of large fluctuations in noise. We report on experimental results obtained with the Aurora-2 connected digit recognition task, and show that the performance for the additive noise cases can be improved by 12.15% over the state-of-the-art multi-condition training if the noise mean is updated every 300 milliseconds.

An integrated approach to feature compensation combining particle filters and hidden Markov models for robust speech recognition

“…Since each cluster encompasses more statistics when compared to the original HMM states, there is a higher possibility that the newly formed probability density function at the cluster level can cover the underlying speech variation to generate appropriate particle filter samples for feature compensation. Additionally, a dynamic joint tracking framework to monitor the clean speech signal and noise simultaneously is also introduced to obtain good noise statistics [13] [14]. In this approach, the information available from clean speech tracking can be effectively used for noise estimation.…”

“…In this chapter, two methods for the estimation of the noise parameters are proposed. The first is based on a particle filter that runs in parallel to the PFC [13]. The second method is based on a MCMC approach [14].…”

“…where x (s) t represents the s th clean speech sample at time t, the noise density is given by N(n (s−1) t , σ n ), and u t = log(e y t −x (s) t − 1) + x t with F(u t ) being the Gaussian cumulative function with mean n (s−1) t and variance σ 2 n . Similarly, to evaluate p(n (s) t |x (s−1) t , y t ), the following relation [13] is used…”