Noise reduction can be achieved relying on something related to spatial and acoustic features: a direction‐of‐arrival (DOA), a fundamental frequency, harmonicity, and so on. It is natural that a suitable noise reduction approach is different depending on the available features, because no all‐purpose noise reduction method has been established under every possible acoustic condition. To achieve noise reduction effectively and efficiently, it is important to explore and acquire well‐suited spatial and acoustic features at all times. Observed acoustic signals originate in some dynamic systems such as a speech production system and behavior of continuous sound source movement, which can be represented by the Markov model in the perspective view, but they are usually distorted by noise and reverberation. In this paper, both feature extraction and feature tracking are attempted by using particle filters aiming at achieving efficient noise reduction. Particle filters can also bring a new feature of the reliability of each DOA estimate based on the effective sample size through the state estimation of spatial features [Mizumachi, Proceedings of the 156th ASA meeting]. In this paper, spatial and acoustic feature extraction by particle filtering is evolved into noise reduction. [Work supported by NEDO, Japan.]