Thomas M. Sullivan,
Multi-Microphone Correlation-Based Processing for Robust Automatic Speech
Recognition, Ph.D Thesis, ECE Department, CMU, August 1996.

Abstract 

Speech recognition systems suffer from degradation in recognition accuracy
when faced with input from noisy and reverberant environments. While most
users prefer a microphone that is placed in the middle of a conference table,
on top of a computer monitor, or mounted in a wall, the recognition accuracy
obtained with such microphones is generally much worse than the accuracy
obtained using a close-talking headset-mounted microphone. Unfortunately,
headset-mounted microphones are often uncomfortable or impractical for users.
Research in recent years on environmental robustness in speech recognition has
concentrated on signal processing using the output of a single microphone to
correct for differences in spectral coloration between microphones used in
the training and testing environments, and to account for the effects of
linear filtering and additive noises present in real testing environments.
This thesis explores the use of microphone arrays to provide further
improvements in speech recognition accuracy.

A novel approach to multiple-microphone processing for the enhancement of
speech input to an automatic speech recognition system is described and
discussed. The system is loosely based on the processing of the binaural
hearing system, but with extensions to an arbitrary number of input
microphones. The processing includes bandpass filtering and nonlinear
rectification of the signals from the microphones to model the effects of
the peripheral auditory system, followed by cross-correlation within each
frequency band of the outputs of the rectifiers from microphone to microphone.
Estimates of the correlated energy within each frequency band are used as the
basis for a feature set for an automatic speech recognition system.

Speech recognition accuracy in natural environments and using
artificially-added noise were compared using the new correlation-based system,
conventional delay-and-sum beamforming, and traditional adaptive filtering
using the Griffiths-Jim algorithm. It was found that the more
computationally-costly correlation-based system provided substantially better
recognition accuracy than previous approaches in pilot experiments using
artificial stimuli, and in experiments using natural speech signals that
were artificially corrupted by additive noise. The correlation-based system
provided a consistent, but much smaller, improvement in recognition accuracy
(relative to previous approaches) for experiments conducted using speech
in two natural environments. It is also demonstrated that the benefit provided
by microphone array processing is complementary to the benefit provided
by single-channel environmental adaptation algorithms such as
codeword-dependent cepstral normalization, regardless of which adaptation
procedure is employed.