Introduction

This paper discusses a novel approach to microphone array signal processing based on a geometrical manipulation on the complex spectrum plane and gives some preliminary experimental results.

Microphone array signal processing is actively studied for various purposes such as improving speech recognition performance in noisy environments[1,2]. The main idea is utilization of differences in path lengths from sources of target and noise signals to multiple microphones.

The simplest technique is Delay-and-Sum (DS) which adjusts delays added to microphone inputs so that the target signal from a particular direction synchronizes across multiple microphones while noises from different directions do not. This technique has an advantage that it requires no training, though it does not give a high performance for noise reduction.

On the other hand, adaptive types of microphone array signal processing such as Griffiths-Jim[3], AMNOR[4], and other adaptive beamforming methods require training the filter coefficients during a silent interval before its operation, though better performance can be obtained compared with DS in such cases. These methods often fail to track rapid changes of environmental characteristics such as moving noise sources, and result in poor improvements in noise reduction even compared with simple DS. Other methods based on blind source separation or independent component analysis assume statistical independence between signal and noise which is not always true.

These methods had mainly aimed at noise cancellation or reduction in the waveform domain. In speech recognition, however, noise reduction in the waveform domain is not necessary; instead, we need noise reduction in feature parameteres such as Mel-Frequency Cepstrum Coefficients (MFCCs) [2].

In this paper, we focus on microphone array signal processing for noise-reduced spectrum estimation for speech recognition.

図 1: Sound propagation to the microphone array

[width=0.8]eps/MicArrayLayout.eps

図 2: Microphone signals $M_i(\omega )$ located on a circle in the complex spectrum plane with the target signal $S(\omega )$ being the circle centroid

[width=0.9]eps/compplanefig2.eps