17-03-2014, 07:18 PM
Abstract
In this paper, a Persian speech recognition system is
proposed to recognize Persian isolated spoken words.
The main contribution of this work in comparison with
the previous ones is simplicity and generality. The
proposed system can be widely used in various real world
applications when the designer does not need so much
expertise in pattern and speech recognition. Due to
generality, robustness, computational simplicity and
reachability, general frequency domain feature as well as
multi-layer percepron neural network as classifier are
considered. To demonstrate the efficiency of the proposed
speech recognition system, a wheeled mobile robot is
navigated in a real domestic environment via Persian
spoken commands. The results indicate the high potential
of the proposed system to deal with real world
applications.
1. Introduction
Speech recognition (SR) is one the most recent fields of
computer science that many researches have been focused
on it during the past decades. As a main motivation, SR
can play the role of a powerful human machine interface.
In other words, SR enables human to reach an old wish,
talking with machines. However this wish will be
completely reached if natural language processing can be
implemented perfectly. The applications of SR are so
much. As a common example, today most of the mobile
phones have a SR subsystem that enables users to
command their phone via their own speech to call to
someone. As a matter of fact, in the near future, SR
systems will take the place of conventional buttons,
switches, handles as well as infra red remote controls
such as TV zapper. Furthermore, SR systems can be
widely used by disabled or ill people when moving in the
environment may be difficult or impossible for them. This
status can be easily found in any hospitals as well as the
house or office of a disabled person. As the state of the
art, in the near future, the computers may be equipped
with the SR system. As a result, it takes the place of
conventional mouse and keyboard. The applications of
SR are not limited to these situations; it can be hardly
used in industry to control the machines via speech. Some
other applications can be easily found in domestics,
security systems, robotics and etc. But designing an
efficient SR system to deal with real world applications is
a difficult and challenging work when several problems
are available. The speech patterns are always full of
uncertainty. Several noises always come from the
environment by other sources. A data acquisition
instruments inherently produce noise. Furthermore, some
perturbations are always available in the own speech
source such as speaker change or even changes in the
voice of a certain person. For example if you say “one”
several times and record your speech signals, you can
find that although you just repeat a certain word, the
signals are different. As a result, the SR system must be
robust enough to deal with these uncertainties. Generally,
SR is a special type of pattern recognition which the
patterns are one-dimensional time domain signals. A
pattern recognition process contains three major stages:
data acquisition and preprocessing, feature extraction and
pattern classification. After data acquisition, due to
prepare the data for feature extraction, several
preprocesses such as filtering must be performed. Since a
suitable feature can improve the robustness and simplify
pattern classification, feature selection must be performed
carefully. As feature extraction is an online process,
computational simplicity is hardly requested. Hence, the
feature must be computationally simple and robust. After
feature selection and extraction, the patterns (features)
must be classified into the desired classes. If the classes
are predefined and determined, supervised learning can be
used. In some cases, the classes are not determined and
the pattern classification algorithm must find the classes
by itself that is called unsupervised learning.
In this paper, a Persian speech recognition system is
proposed to recognize Persian isolated spoken words.
The main contribution of this work in comparison with
the previous ones is simplicity and generality. The
proposed system can be widely used in various real world
applications when the designer does not need so much
expertise in pattern and speech recognition. Due to
generality, robustness, computational simplicity and
reachability, general frequency domain feature as well as
multi-layer percepron neural network as classifier are
considered. To demonstrate the efficiency of the proposed
speech recognition system, a wheeled mobile robot is
navigated in a real domestic environment via Persian
spoken commands. The results indicate the high potential
of the proposed system to deal with real world
applications.
1. Introduction
Speech recognition (SR) is one the most recent fields of
computer science that many researches have been focused
on it during the past decades. As a main motivation, SR
can play the role of a powerful human machine interface.
In other words, SR enables human to reach an old wish,
talking with machines. However this wish will be
completely reached if natural language processing can be
implemented perfectly. The applications of SR are so
much. As a common example, today most of the mobile
phones have a SR subsystem that enables users to
command their phone via their own speech to call to
someone. As a matter of fact, in the near future, SR
systems will take the place of conventional buttons,
switches, handles as well as infra red remote controls
such as TV zapper. Furthermore, SR systems can be
widely used by disabled or ill people when moving in the
environment may be difficult or impossible for them. This
status can be easily found in any hospitals as well as the
house or office of a disabled person. As the state of the
art, in the near future, the computers may be equipped
with the SR system. As a result, it takes the place of
conventional mouse and keyboard. The applications of
SR are not limited to these situations; it can be hardly
used in industry to control the machines via speech. Some
other applications can be easily found in domestics,
security systems, robotics and etc. But designing an
efficient SR system to deal with real world applications is
a difficult and challenging work when several problems
are available. The speech patterns are always full of
uncertainty. Several noises always come from the
environment by other sources. A data acquisition
instruments inherently produce noise. Furthermore, some
perturbations are always available in the own speech
source such as speaker change or even changes in the
voice of a certain person. For example if you say “one”
several times and record your speech signals, you can
find that although you just repeat a certain word, the
signals are different. As a result, the SR system must be
robust enough to deal with these uncertainties. Generally,
SR is a special type of pattern recognition which the
patterns are one-dimensional time domain signals. A
pattern recognition process contains three major stages:
data acquisition and preprocessing, feature extraction and
pattern classification. After data acquisition, due to
prepare the data for feature extraction, several
preprocesses such as filtering must be performed. Since a
suitable feature can improve the robustness and simplify
pattern classification, feature selection must be performed
carefully. As feature extraction is an online process,
computational simplicity is hardly requested. Hence, the
feature must be computationally simple and robust. After
feature selection and extraction, the patterns (features)
must be classified into the desired classes. If the classes
are predefined and determined, supervised learning can be
used. In some cases, the classes are not determined and
the pattern classification algorithm must find the classes
by itself that is called unsupervised learning.