Abstract: Speech to text in Malay language is a system that converts Malay speech into text. The Malay language recognition system is still limited, thus, this paper aims to investigate the performance of ten Malay words obtained from the online Malay news. The methodology consists of three stages, which are preprocessing, feature extraction, and speech classification. In preprocessing stage, the speech samples are filtered using pre emphasis. After that, feature extraction method is applied to the samples using Mel Frequency Cepstrum Coefficient (MFCC). Lastly, speech classification is performed using Feedforward Neural Network (FFNN). The accuracy of the classification is further investigated based on the hidden layer size. From experimentation, the classifier with 40 hidden neurons shows the highest classification rate which is 94%.
Abstract: Statement of the automatic speech recognition
problem, the assignment of speech recognition and the application
fields are shown in the paper. At the same time as Azerbaijan speech,
the establishment principles of speech recognition system and the
problems arising in the system are investigated. The computing algorithms of speech features, being the main part
of speech recognition system, are analyzed. From this point of view,
the determination algorithms of Mel Frequency Cepstral Coefficients
(MFCC) and Linear Predictive Coding (LPC) coefficients expressing
the basic speech features are developed. Combined use of cepstrals of
MFCC and LPC in speech recognition system is suggested to
improve the reliability of speech recognition system. To this end, the
recognition system is divided into MFCC and LPC-based recognition
subsystems. The training and recognition processes are realized in
both subsystems separately, and recognition system gets the decision
being the same results of each subsystems. This results in decrease of
error rate during recognition. The training and recognition processes are realized by artificial
neural networks in the automatic speech recognition system. The
neural networks are trained by the conjugate gradient method. In the
paper the problems observed by the number of speech features at
training the neural networks of MFCC and LPC-based speech
recognition subsystems are investigated. The variety of results of neural networks trained from different
initial points in training process is analyzed. Methodology of
combined use of neural networks trained from different initial points
in speech recognition system is suggested to improve the reliability
of recognition system and increase the recognition quality, and
obtained practical results are shown.
Abstract: Speech recognition is of an important contribution in promoting new technologies in human computer interaction. Today, there is a growing need to employ speech technology in daily life and business activities. However, speech recognition is a challenging task that requires different stages before obtaining the desired output. Among automatic speech recognition (ASR) components is the feature extraction process, which parameterizes the speech signal to produce the corresponding feature vectors. Feature extraction process aims at approximating the linguistic content that is conveyed by the input speech signal. In speech processing field, there are several methods to extract speech features, however, Mel Frequency Cepstral Coefficients (MFCC) is the popular technique. It has been long observed that the MFCC is dominantly used in the well-known recognizers such as the Carnegie Mellon University (CMU) Sphinx and the Markov Model Toolkit (HTK). Hence, this paper focuses on the MFCC method as the standard choice to identify the different speech segments in order to obtain the language phonemes for further training and decoding steps. Due to MFCC good performance, the previous studies show that the MFCC dominates the Arabic ASR research. In this paper, we demonstrate MFCC as well as the intermediate steps that are performed to get these coefficients using the HTK toolkit.
Abstract: We propose a system to real environmental noise and
channel mismatch for forensic speaker verification systems. This
method is based on suppressing various types of real environmental
noise by using independent component analysis (ICA) algorithm.
The enhanced speech signal is applied to mel frequency cepstral
coefficients (MFCC) or MFCC feature warping to extract the
essential characteristics of the speech signal. Channel effects are
reduced using an intermediate vector (i-vector) and probabilistic
linear discriminant analysis (PLDA) approach for classification. The
proposed algorithm is evaluated by using an Australian forensic voice
comparison database, combined with car, street and home noises
from QUT-NOISE at a signal to noise ratio (SNR) ranging from -10
dB to 10 dB. Experimental results indicate that the MFCC feature
warping-ICA achieves a reduction in equal error rate about (48.22%,
44.66%, and 50.07%) over using MFCC feature warping when the
test speech signals are corrupted with random sessions of street, car,
and home noises at -10 dB SNR.
Abstract: Speaker Identification (SI) is the task of establishing
identity of an individual based on his/her voice characteristics. The SI
task is typically achieved by two-stage signal processing: training and
testing. The training process calculates speaker specific feature
parameters from the speech and generates speaker models
accordingly. In the testing phase, speech samples from unknown
speakers are compared with the models and classified. Even though
performance of speaker identification systems has improved due to
recent advances in speech processing techniques, there is still need of
improvement. In this paper, a Closed-Set Tex-Independent Speaker
Identification System (CISI) based on a Multiple Classifier System
(MCS) is proposed, using Mel Frequency Cepstrum Coefficient
(MFCC) as feature extraction and suitable combination of vector
quantization (VQ) and Gaussian Mixture Model (GMM) together
with Expectation Maximization algorithm (EM) for speaker
modeling. The use of Voice Activity Detector (VAD) with a hybrid
approach based on Short Time Energy (STE) and Statistical
Modeling of Background Noise in the pre-processing step of the
feature extraction yields a better and more robust automatic speaker
identification system. Also investigation of Linde-Buzo-Gray (LBG)
clustering algorithm for initialization of GMM, for estimating the
underlying parameters, in the EM step improved the convergence rate
and systems performance. It also uses relative index as confidence
measures in case of contradiction in identification process by GMM
and VQ as well. Simulation results carried out on voxforge.org
speech database using MATLAB highlight the efficacy of the
proposed method compared to earlier work.
Abstract: Over the past few years, the online multimedia
collection has grown at a fast pace. Several companies showed
interest to study the different ways to organise the amount of audio
information without the need of human intervention to generate
metadata. In the past few years, many applications have emerged on
the market which are capable of identifying a piece of music in a
short time. Different audio effects and degradation make it much
harder to identify the unknown piece. In this paper, an audio
fingerprinting system which makes use of a non-parametric based
algorithm is presented. Parametric analysis is also performed using
Gaussian Mixture Models (GMMs). The feature extraction methods
employed are the Mel Spectrum Coefficients and the MPEG-7 basic
descriptors. Bin numbers replaced the extracted feature coefficients
during the non-parametric modelling. The results show that nonparametric
analysis offer potential results as the ones mentioned in
the literature.
Abstract: The performance and analysis of speech recognition
system is illustrated in this paper. An approach to recognize the
English word corresponding to digit (0-9) spoken by 2 different
speakers is captured in noise free environment. For feature extraction,
speech Mel frequency cepstral coefficients (MFCC) has been used
which gives a set of feature vectors from recorded speech samples.
Neural network model is used to enhance the recognition
performance. Feed forward neural network with back propagation
algorithm model is used. However other speech recognition
techniques such as HMM, DTW exist. All experiments are carried
out on Matlab.
Abstract: Paper presents an comparative evaluation of features extraction algorithm for a real-time isolated word recognition system
based on FPGA. The Mel-frequency cepstral, linear frequency cepstral, linear predictive and their cepstral coefficients were
implemented in hardware/software design. The proposed system was investigated in speaker dependent mode for 100 different
Lithuanian words. The robustness of features extraction algorithms was tested recognizing the speech records at different signal to noise rates. The experiments on clean records show highest accuracy for Mel-frequency cepstral and linear frequency cepstral coefficients. For records with 15 dB signal to noise rate the linear predictive cepstral coefficients gives best result. The hard and soft part of the system is clocked on 50 MHz and 100 MHz accordingly. For the classification purpose the pipelined dynamic time warping core was implemented. The proposed word recognition system satisfy the real-time requirements and is suitable for applications in embedded systems.
Abstract: The goal of speech parameterization is to extract the relevant information about what is being spoken from the audio signal. In speech recognition systems Mel-Frequency Cepstral Coefficients (MFCC) and Relative Spectral Mel-Frequency Cepstral Coefficients (RASTA-MFCC) are the two main techniques used. It will be shown in this paper that it presents some modifications to the original MFCC method. In our work the effectiveness of proposed changes to MFCC called Modified Function Cepstral Coefficients (MODFCC) were tested and compared against the original MFCC and RASTA-MFCC features. The prosodic features such as jitter and shimmer are added to baseline spectral features. The above-mentioned techniques were tested with impulsive signals under various noisy conditions within AURORA databases.
Abstract: In this paper, we propose effective system for digital music retrieval. We divided proposed system into Client and Server. Client part consists of pre-processing and Content-based feature extraction stages. In pre-processing stage, we minimized Time code Gap that is occurred among same music contents. As content-based feature, first-order differentiated MFCC were used. These presented approximately envelop of music feature sequences. Server part included Music Server and Music Matching stage. Extracted features from 1,000 digital music files were stored in Music Server. In Music Matching stage, we found retrieval result through similarity measure by DTW. In experiment, we used 450 queries. These were made by mixing different compression standards and sound qualities from 50 digital music files. Retrieval accurate indicated 97% and retrieval time was average 15ms in every single query. Out experiment proved that proposed system is effective in retrieve digital music and robust at various user environments of web.
Abstract: Real world Speaker Identification (SI) application
differs from ideal or laboratory conditions causing perturbations that
leads to a mismatch between the training and testing environment
and degrade the performance drastically. Many strategies have been
adopted to cope with acoustical degradation; wavelet based Bayesian
marginal model is one of them. But Bayesian marginal models
cannot model the inter-scale statistical dependencies of different
wavelet scales. Simple nonlinear estimators for wavelet based
denoising assume that the wavelet coefficients in different scales are
independent in nature. However wavelet coefficients have significant
inter-scale dependency. This paper enhances this inter-scale
dependency property by a Circularly Symmetric Probability Density
Function (CS-PDF) related to the family of Spherically Invariant
Random Processes (SIRPs) in Log Gabor Wavelet (LGW) domain
and corresponding joint shrinkage estimator is derived by Maximum
a Posteriori (MAP) estimator. A framework is proposed based on
these to denoise speech signal for automatic speaker identification
problems. The robustness of the proposed framework is tested for
Text Independent Speaker Identification application on 100 speakers
of POLYCOST and 100 speakers of YOHO speech database in three
different noise environments. Experimental results show that the
proposed estimator yields a higher improvement in identification
accuracy compared to other estimators on popular Gaussian Mixture
Model (GMM) based speaker model and Mel-Frequency Cepstral
Coefficient (MFCC) features.
Abstract: This paper presents an ESN-based Arabic phoneme
recognition system trained with supervised, forced and combined
supervised/forced supervised learning algorithms. Mel-Frequency
Cepstrum Coefficients (MFCCs) and Linear Predictive Code (LPC)
techniques are used and compared as the input feature extraction
technique. The system is evaluated using 6 speakers from the King
Abdulaziz Arabic Phonetics Database (KAPD) for Saudi Arabia
dialectic and 34 speakers from the Center for Spoken Language
Understanding (CSLU2002) database of speakers with different
dialectics from 12 Arabic countries. Results for the KAPD and
CSLU2002 Arabic databases show phoneme recognition
performances of 72.31% and 38.20% respectively.
Abstract: In this paper, an algorithm for detecting and attenuating
puff noises frequently generated under the mobile environment is
proposed. As a baseline system, puff detection system is designed
based on Gaussian Mixture Model (GMM), and 39th Mel Frequency
Cepstral Coefficient (MFCC) is extracted as feature parameters. To
improve the detection performance, effective acoustic features for puff
detection are proposed. In addition, detected puff intervals are
attenuated by high-pass filtering. The speech recognition rate was
measured for evaluation and confusion matrix and ROC curve are used
to confirm the validity of the proposed system.
Abstract: Despite the fact that Arabic language is currently one
of the most common languages worldwide, there has been only a
little research on Arabic speech recognition relative to other
languages such as English and Japanese. Generally, digital speech
processing and voice recognition algorithms are of special
importance for designing efficient, accurate, as well as fast automatic
speech recognition systems. However, the speech recognition process
carried out in this paper is divided into three stages as follows: firstly,
the signal is preprocessed to reduce noise effects. After that, the
signal is digitized and hearingized. Consequently, the voice activity
regions are segmented using voice activity detection (VAD)
algorithm. Secondly, features are extracted from the speech signal
using Mel-frequency cepstral coefficients (MFCC) algorithm.
Moreover, delta and acceleration (delta-delta) coefficients have been
added for the reason of improving the recognition accuracy. Finally,
each test word-s features are compared to the training database using
dynamic time warping (DTW) algorithm. Utilizing the best set up
made for all affected parameters to the aforementioned techniques,
the proposed system achieved a recognition rate of about 98.5%
which outperformed other HMM and ANN-based approaches
available in the literature.
Abstract: Automatic detection of syllable repetition is one of the
important parameter in assessing the stuttered speech objectively.
The existing method which uses artificial neural network (ANN)
requires high levels of agreement as prerequisite before attempting to
train and test ANNs to separate fluent and nonfluent. We propose
automatic detection method for syllable repetition in read speech for
objective assessment of stuttered disfluencies which uses a novel
approach and has four stages comprising of segmentation, feature
extraction, score matching and decision logic. Feature extraction is
implemented using well know Mel frequency Cepstra coefficient
(MFCC). Score matching is done using Dynamic Time Warping
(DTW) between the syllables. The Decision logic is implemented by
Perceptron based on the score given by score matching. Although
many methods are available for segmentation, in this paper it is done
manually. Here the assessment by human judges on the read speech
of 10 adults who stutter are described using corresponding method
and the result was 83%.
Abstract: A state of the art Speaker Identification (SI) system
requires a robust feature extraction unit followed by a speaker
modeling scheme for generalized representation of these features.
Over the years, Mel-Frequency Cepstral Coefficients (MFCC)
modeled on the human auditory system has been used as a standard
acoustic feature set for speech related applications. On a recent
contribution by authors, it has been shown that the Inverted Mel-
Frequency Cepstral Coefficients (IMFCC) is useful feature set for
SI, which contains complementary information present in high
frequency region. This paper introduces the Gaussian shaped filter
(GF) while calculating MFCC and IMFCC in place of typical
triangular shaped bins. The objective is to introduce a higher
amount of correlation between subband outputs. The performances
of both MFCC & IMFCC improve with GF over conventional
triangular filter (TF) based implementation, individually as well as
in combination. With GMM as speaker modeling paradigm, the
performances of proposed GF based MFCC and IMFCC in
individual and fused mode have been verified in two standard
databases YOHO, (Microphone Speech) and POLYCOST
(Telephone Speech) each of which has more than 130 speakers.
Abstract: In this paper, we propose a practical digital music matching system that is robust to variation in sound qualities. The proposed system is subdivided into two parts: client and server. The client part consists of the input, preprocessing and feature extraction modules. The preprocessing module, including the music onset module, revises the value gap occurring on the time axis between identical songs of different formats. The proposed method uses delta-grouped Mel frequency cepstral coefficients (MFCCs) to extract music features that are robust to changes in sound quality. According to the number of sound quality formats (SQFs) used, a music server is constructed with a feature database (FD) that contains different sub feature databases (SFDs). When the proposed system receives a music file, the selection module selects an appropriate SFD from a feature database; the selected SFD is subsequently used by the matching module. In this study, we used 3,000 queries for matching experiments in three cases with different FDs. In each case, we used 1,000 queries constructed by mixing 8 SQFs and 125 songs. The success rate of music matching improved from 88.6% when using single a single SFD to 93.2% when using quadruple SFDs. By this experiment, we proved that the proposed method is robust to various sound qualities.
Abstract: An emotional speech recognition system for the
applications on smart phones was proposed in this study to combine
with 3G mobile communications and social networks to provide users
and their groups with more interaction and care. This study developed
a mechanism using the support vector machines (SVM) to recognize
the emotions of speech such as happiness, anger, sadness and normal.
The mechanism uses a hierarchical classifier to adjust the weights of
acoustic features and divides various parameters into the categories of
energy and frequency for training. In this study, 28 commonly used
acoustic features including pitch and volume were proposed for
training. In addition, a time-frequency parameter obtained by
continuous wavelet transforms was also used to identify the accent and
intonation in a sentence during the recognition process. The Berlin
Database of Emotional Speech was used by dividing the speech into
male and female data sets for training. According to the experimental
results, the accuracies of male and female test sets were increased by
4.6% and 5.2% respectively after using the time-frequency parameter
for classifying happy and angry emotions. For the classification of all
emotions, the average accuracy, including male and female data, was
63.5% for the test set and 90.9% for the whole data set.
Abstract: Heart sound is an acoustic signal and many techniques
used nowadays for human recognition tasks borrow speech recognition
techniques. One popular choice for feature extraction of accoustic
signals is the Mel Frequency Cepstral Coefficients (MFCC) which
maps the signal onto a non-linear Mel-Scale that mimics the human
hearing. However the Mel-Scale is almost linear in the frequency
region of heart sounds and thus should produce similar results with
the standard cepstral coefficients (CC). In this paper, MFCC is
investigated to see if it produces superior results for PCG based
human identification system compared to CC. Results show that the
MFCC system is still superior to CC despite linear filter-banks in
the lower frequency range, giving up to 95% correct recognition rate
for MFCC and 90% for CC. Further experiments show that the high
recognition rate is due to the implementation of filter-banks and not
from Mel-Scaling.
Abstract: This paper proposes evaluation of sound parameterization methods in recognizing some spoken Arabic words, namely digits from zero to nine. Each isolated spoken word is represented by a single template based on a specific recognition feature, and the recognition is based on the Euclidean distance from those templates. The performance analysis of recognition is based on four parameterization features: the Burg Spectrum Analysis, the Walsh Spectrum Analysis, the Thomson Multitaper Spectrum Analysis and the Mel Frequency Cepstral Coefficients (MFCC) features. The main aim of this paper was to compare, analyze, and discuss the outcomes of spoken Arabic digits recognition systems based on the selected recognition features. The results acqired confirm that the use of MFCC features is a very promising method in recognizing Spoken Arabic digits.