Switching Convolutional Beamformer

Tomohiro Nakatani, Rintaro Ikeshita, Naoyuki Kamo, Keisuke Kinoshita, Shoko Araki, Hiroshi Sawada

Abstract: This paper proposes a time-varying Convolutional BeamFormer (CBF), called a switching CBF, which can capture time-varying characteristics of an observed signal to perform beamforming and dereverberation accurately and simultaneously. With a switching CBF, time frames of a time-varying observed signal are grouped into several clusters, each of which can be taken as time-invariant, and individual clusters are separately processed by different time-invariant CBFs. Conventionally, a switching BeamFormer (BF) and a switching Weighted Prediction Error (WPE) dereverberation filter have been shown effective for the respective problems. This paper presents a method to integrate and jointly optimize them based on Maximum Likelihood (ML) estimation and extends it to work with a Neural Network (NN)-based spectral prior based on Maximum a Posteriori (MAP) estimation. Experiments show that a switching CBF largely outperforms a conventional time-invariant CBF in terms of improved ASR scores.