17-03-2014, 10:40 AM
Abstract - One of the best current methods for modeling
dynamic speech signal is using of HMM model. The speech
recognition systems based on HMM can be able to compute
the best likelihood measure between unknown input pattern
and reference models by using Viterbi algorithm. Whereas
such algorithm is based on dynamic programming, it
consists of many computations with increasing number of
reference words. In this paper, we will present a new
evolutionary methodology based on synergic HMM and GA
that will be able to compute likelihood measurement
between unknown input pattern and reference patterns in
the parallel form and based on cellular automata. We
introduce this algorithm as HGC. The HGC algorithm will
be compared with the Viterbi algorithm from the
“recognition accuracy” and “recognition speed” viewpoints.
Obtained results show that the HGC and Viterbi algorithms
are close from “recognition accuracy” viewpoint, but HGC
isso faster than the Viterbi
I. INTRODUCTION
History of speech recognition with machines refers to
several years ago. Block diagram of a statistical speech
recognizer system has been shown in figure (1). The purpose of the front end processing stage is to
parameterize the incoming speech signal. The reason for
this is two fold: firstly, to represent the signal in a more
compact form and secondly, to extract relevant acoustic
features from the speech signal to be used in the
recognition process.
Each statistical speech recognizer system operate in one of
two modes: Training mode and recognition mode. In either
mode, the front end analysis module accepts the input
speech utterance and divides it into blocks of equal
numbers of samples called frames. Hence, a sequence of
feature measurements are performed on each frame and
the acoustic information is extracted from the frames. The
extracted information forms a sequence of feature vectors
which can be used to characterize the input “speech
patterns”. In the training mode, the output speech patterns
from the front end analysis module are used for creating
models .In the recognition mode, the output speech pattern
of front end analysis module is called the “Observation
Sequence” [1]. In this mode, likelihood measurement
between the observation sequence and the reference
models via pattern matching module will be computed.
While the Hidden Markov Model can be able to model
dynamics of speech signal in statistical form, it has been
known as the best model for speech recognition [1, 2].
Also, some of other methods such as Dynamic Time
Warping (DTW) [5] and Neural Networks [6] have been
applied for speech recognition, but the performance of
none of them is as well as performance of speech
recognition systems based on HMM. The HMM - by using
of an algorithm that is called Viterbi algorithm - is able to
compute likelihood measurement between observation
sequence and reference patterns. Since Viterbi algorithm is
based on dynamic programming, it involves many
computations. In order to accelerate such algorithm,
several software and hardware solutions have been
presented. Beam and Histogram searches [7, 8] are two
types of the best software solutions. Also several special
hardware with parallel and pipeline processing capabilities
were designed [9, 10, and 11]. In this paper, a stochastic
method based on synergic HMM and GA will be
suggested that can compute likelihood measurement
between observation sequence and reference pattern
instead of using Viterbi algorithm. The proposed method
is called HGC. Results of experiments show that HGC
algorithm has better performance than Viterbi algorithm.
dynamic speech signal is using of HMM model. The speech
recognition systems based on HMM can be able to compute
the best likelihood measure between unknown input pattern
and reference models by using Viterbi algorithm. Whereas
such algorithm is based on dynamic programming, it
consists of many computations with increasing number of
reference words. In this paper, we will present a new
evolutionary methodology based on synergic HMM and GA
that will be able to compute likelihood measurement
between unknown input pattern and reference patterns in
the parallel form and based on cellular automata. We
introduce this algorithm as HGC. The HGC algorithm will
be compared with the Viterbi algorithm from the
“recognition accuracy” and “recognition speed” viewpoints.
Obtained results show that the HGC and Viterbi algorithms
are close from “recognition accuracy” viewpoint, but HGC
isso faster than the Viterbi
I. INTRODUCTION
History of speech recognition with machines refers to
several years ago. Block diagram of a statistical speech
recognizer system has been shown in figure (1). The purpose of the front end processing stage is to
parameterize the incoming speech signal. The reason for
this is two fold: firstly, to represent the signal in a more
compact form and secondly, to extract relevant acoustic
features from the speech signal to be used in the
recognition process.
Each statistical speech recognizer system operate in one of
two modes: Training mode and recognition mode. In either
mode, the front end analysis module accepts the input
speech utterance and divides it into blocks of equal
numbers of samples called frames. Hence, a sequence of
feature measurements are performed on each frame and
the acoustic information is extracted from the frames. The
extracted information forms a sequence of feature vectors
which can be used to characterize the input “speech
patterns”. In the training mode, the output speech patterns
from the front end analysis module are used for creating
models .In the recognition mode, the output speech pattern
of front end analysis module is called the “Observation
Sequence” [1]. In this mode, likelihood measurement
between the observation sequence and the reference
models via pattern matching module will be computed.
While the Hidden Markov Model can be able to model
dynamics of speech signal in statistical form, it has been
known as the best model for speech recognition [1, 2].
Also, some of other methods such as Dynamic Time
Warping (DTW) [5] and Neural Networks [6] have been
applied for speech recognition, but the performance of
none of them is as well as performance of speech
recognition systems based on HMM. The HMM - by using
of an algorithm that is called Viterbi algorithm - is able to
compute likelihood measurement between observation
sequence and reference patterns. Since Viterbi algorithm is
based on dynamic programming, it involves many
computations. In order to accelerate such algorithm,
several software and hardware solutions have been
presented. Beam and Histogram searches [7, 8] are two
types of the best software solutions. Also several special
hardware with parallel and pipeline processing capabilities
were designed [9, 10, and 11]. In this paper, a stochastic
method based on synergic HMM and GA will be
suggested that can compute likelihood measurement
between observation sequence and reference pattern
instead of using Viterbi algorithm. The proposed method
is called HGC. Results of experiments show that HGC
algorithm has better performance than Viterbi algorithm.