Abstract:
Revcrberation is one of thc primary factors that dcgrade the quality of speech
signals whcn recorded by a distant microphonc in order to facilitate hands-frce
communication. Undoing the effect of reverberation is still a challenging problem
especially when additive noise and time-varying acoustic channels arc considered. In
this dissertation, several multimicrophone dcreverbcration techniqucs are developed
that can dereverberatc the rccordcd specch as well as improve the signal-ta-noise
ratio (SNR) considering a practical acoustic environment. The methods are based
on the adaptive estimation of the long acoustic impulse responses (AIRs) using the
multichannel LMS (MCLMS) algorithm. Although the MCLMS algorithm is attractive
for its simplicity and computational efficiency, it suffers from slow convergence rate,
step-size ambiguity, and last but not the least, lack of robustness in the presence
of noise. A variable-step-sizc frequency-domain MCLMS algorithm is proposcd that
can ensure stability and optimal convergence speed both in the noise-frcc and noisy
conditions. To improvc the noisc-robustness of thc class of MCLMS algorithms, two
novcl solutions, namely, excitation-driven MCLMS and spcctrally constraint MCLMS
algorithms are proposed that can successfully estimate thc long AIRs with reasonable
accuracy.
Based on adaptive cstimation of the AIRs, two different dereverberation techniques
are proposed. In the first approach, derevcrberation is achieved by suppressing the
late reverberation using channel shortening technique and the SNR is improved by
delay-and-sum beamforming. The proposed shortening algorithm is also optimized
so that it makes a trade-off between shortening performance and spectral distortion
in the dcreverberated speech. In the sccond approach, the power of the speech
components in the receivcd microphone signals are first enhanced by an eigenfilter
and then a block-adaptive zero-forcing equalizer is employed to eliminate the channel
distortion introduccd by the AIRs and cigenfilter. The cigenfilter is cfficiently estimated
avoiding the tedious Cholesky factorization and it also resists spectral nulls so that
noise amplification is mitigated at the output of the zero-forcing equalizer. Extensive
experiments are conducted, using both simulated and real reverberant acoustic
channels, which demonstrate that the proposed methods can offer better speech quality
and SNR improvement as compared to the state-of-the-art dcreverberation techniques.
