Noise Robust Automatic Speech Recognition

The recent advancements in automatic speech
recognition (ASR)
systems have expedited their commercial deployment.
Still, a major
challenge faced by the current ASR systems is their poor
performance for any mismatch between training and
testing
conditions. This book presents an in-depth analysis
of multi-
stream combination approach to improve the robustness
of ASR
systems for mismatch caused by additive noise. The
book also
discusses that multi-stream approach is complementary
to other
approaches for noise robustness, and how they can be
used
together. The selection of feature streams and
allocation of weights
to each feature stream are two important issues in
multi-stream.
Both the issues are analyzed, and a step-wise
approach is
presented to realize a significant improvement in
performance by a
multi-stream system compared to that of a
state-of-the-art
system. The analysis proposed in this
work can also be applied to other pattern recognition
tasks such as
speaker recogntion and text-processing, rendering
this book useful
for a wide audience ranging from beginner to advance
stage users
working in ASR as well as other pattern recognition
domains.

Autorentext

Dr. Hemant Misra is an active researcher in the areas of
automatic speech/speaker
recognition, text processing and machine learning, and is a
post-doc researcher at
University of Glasgow. His academic and industry research career
spans more than a
decade. He completed M.S. and Ph.D. from IIT Madras, India, and
EPFL, Switzerland,
respectively.

Klappentext

The recent advancements in automatic speech
recognition (ASR)
systems have expedited their commercial deployment.
Still, a major
challenge faced by the current ASR systems is their poor
performance for any mismatch between training and
testing
conditions. This book presents an in-depth analysis
of multi-
stream combination approach to improve the robustness
of ASR
systems for mismatch caused by additive noise. The
book also
discusses that multi-stream approach is complementary
to other
approaches for noise robustness, and how they can be
used
together. The selection of feature streams and
allocation of weights
to each feature stream are two important issues in
multi-stream.
Both the issues are analyzed, and a step-wise
approach is
presented to realize a significant improvement in
performance by a
multi-stream system compared to that of a
state-of-the-art
system. The analysis proposed in this
work can also be applied to other pattern recognition
tasks such as
speaker recogntion and text-processing, rendering
this book useful
for a wide audience ranging from beginner to advance
stage users
working in ASR as well as other pattern recognition
domains.