Subband particle filtering for speech enhancement

Particle filters have recently been applied to speech enhancement when the input speech signal is modeled as a time-varying autoregressive process with stochastically evolving parameters. This type of modeling results in a nonlinear and conditionally Gaussian state-space system that is not amenable to analytical solutions. Prior work in this area involved signal processing in the fullband domain and assumed white Gaussian noise with known variance. This paper extends such ideas to subband domain particle filters and colored noise. Experimental results indicate that the subband particle filter achieves higher segmental SNR than the fullband algorithm and is effective in dealing with colored noise without increasing the computational complexity.