Identification of frequency domain using quantum based optimization neural networks

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Special Issue: 07 | May-2014, Available @ http://www.ijret.org 443
IDENTIFICATION OF FREQUENCY DOMAIN USING QUANTUM
BASED OPTIMIZATION NEURAL NETWORKS
Dhivya bharathi1
, T.Karthikeyan2
, P.Hemalatha3
, P. Joy kinshy4
1
PG Scholar, Computer Science and Engineering, Knowledge Institute of Technology, Salem
2
Assistant Professor, Computer Science and Engineering, Knowledge Institute of Technology, Salem
3
4
Abstract
Voice Conversion (VC) is a zone of speech processing that contract with the conversion of the apparent speaker identity. This paper
presents a new conversion method for text to speech and Voice to voice conversions. Text to Speech conversion uses phonematic
concatenation for conversion and Voice conversion uses MLP Quantum Neural Network for transformations. The proposed method
consists of two stages. In first stage, Features extraction of LSF and pitch residual based data from source and target speaker. In
second stage where we use MLP Quantum neural network which is trained to learn the nonlinear mapping function for source to
target speech transformation using the features extracted in the first phase. The proposed method can efficiently increase the quality
and lack of naturalness of the converted speech. The proposed model is tested by male and female speakers with average duration.
Key Words: Voice Conversion, Line Spectral Pairs, Quantum Neural Networks, Multi-layer Perceptron’s, Phonematic
Concatenation, TTS, and GMM.
---------------------------------------------------------------------***----------------------------------------------------------------------
1. INTRODUCTION
Text-to-speech (TTS) convention transforms linguistic
information stored as data or text into voice. It is far and wide
used in acoustic understanding devices for blind people.
Potential applications of High Quality TTS Systems are in-
deed numerous. Here are some examples:
Telecommunications services, Multimedia, man-machine
communication, Vocal Monitoring, Talking books and toys. A
voice conversion system modifies the utterance of a source
speaker to be perceived as spoken by a specified target
speaker. While the speaker identity is transformed, the speech
content (linguistic information) is left unaltered.
There are many applications of voice conversion including
customizing voices for TTS systems, transforming voice-overs
in adverts and films to sound like that of a well-known
celebrity, enhancing the speech of impaired speakers,
international dubbing, health-care, multi-media, language
education, voice restoration in old documents/movies, to
create a speech database in a cost-efficient manner in the field
of speech synthesis and to dub character voices in television
programs. Generally, VC operation can be divided into two
parts, vocal tract parameter conversion and excitation signal
conversion. It is believed that speaker characteristics are
inherited in the vocal tract parameter. One of the most widely
used spectrum conversion methods is statistical parametric
approach, Gaussian Mixture Model (GMM) based algorithm.
From the other point of view, another transformation paradigm
was also conducted, named frequency warping.
This transformation function maps significant positions of the
frequency axis (e.g., central frequency of formants) from the
source speaker to the target speaker. As this method does not
modify the fine spectral details of the source spectrum, it
preserves very well the quality of the converted speech.
However, the accuracy was less than that of GMM-based VC.
Another interest growing in voice conversion field is by using
Quantum Neural Networks (QNNs).This interest network will
perform nonlinear mapping.
Voice transformation should be performed from source speech
to target speech without losing or modifying the original
speech content.VC is performed in two phases: the first one is
a training phase, in this phase the speech features of both
source and target speaker are extracted and appropriate
mapping rules are generated for transforming the parameters
of the source speaker onto those of the target speaker. In the
second testing phase, the mapping rules developed in the
training stage are used to transform the features of source
voice signal in order to possess the characteristics of the target
voice. The vocal tract and prosodic parameters are modified
using signal processing algorithms. There are many
applications where intra gender and inter gender voice
conversion is required.

__________________________________________________________________________________________
2. PREVIOUS WORK
In voice conversion, the quality and intelligibility of the
Synthetic speech depends on time and spectral expansion,
compression, pitch modifications, the glottal excitation, shape
and also on the reproduction. Gaussian mixture models
(GMMs) is used to partition the acoustic space of the source
speaker into overlapping classes called as soft partitions.
Using these soft partitions, a continuous probabilistic linear
transformation function for vector is obtained. This
transformation function contained parametric representation of
the spectral envelope and modified the GMM approach in [4].
However, the quality and naturalness of the converted speech
signal is found inadequate due to reconstruction of speech
signal using the large number of parameters and it results into
over smoothing.
Researchers have also provided the solution for over
smoothing problem like hybridization of GMM with
frequency warping [5] and GMM with codebook mapping.
De-sai et.al. [6] have compared the performance of the ANN
and GMM and it is reported that the ANN performs better than
GMM. it is reported that the structural GMM performs better
than GMM. GMM with GA based method [5][6] has been
developed using LSP. The conversion function has been
proposed using the probabilistic model based on inter-speaker
dependencies and cross correlation probabilities between the
source and target speaker .Chen et.al have pro-posed improved
method [6][7] for VC, in which the input frames are divided
into voiced and unvoiced frames. Chen et.al has proposed
improved method [4] for VC, in which the input frames are
divided into voiced and unvoiced frames.
The voiced frames are mapped from source to target t using
GMM and unvoiced frames are stretched or compressed for
conversion based on the ratio of vocal tract length (VTL) of
source to target. Artificial neural network has been used for
VC to exploit the nonlinear relationship between the vocal
tract shape of source and target speaker [4].Line spectral pitch
has been used for VC where LSP and fundamental frequency
(Fo) are used to extract the source and target features of
parallel set of data. Using these features the RBF based neural
network is trained for an appropriate mapping function that
transforms the vocal spectral and glottal excitation cues of the
source speaker in to target speaker’s acoustic space [3].
3. PROPOSED METHOD
The proposed algorithm consists of two phases. In first phase,
we extract the Line Spectral Frequency and pitch residual
based features from source and target speaker data. It is
followed by the second phase where we use MLP neural
network and GMM which is trained to learn the nonlinear
mapping function for source to target speech transformation
using the features extracted in the first phase.
3.1 Text-To-Speech Conversion by Phonemic
Concatenation
Text to speech conversion is as follows:-
 GUI implementation of the Mat lab Module.
 Recording of voice samples through microphone.
 Phoneme extraction by the use of spectrogram.
 Concatenation of phonemes to create any desired word.
 Comparison of concatenated word with the original
word. [2]
3.2 LSF, Pitch Residual Based Voice Conversion
When the LSF are in ascending order in the range [0, 1], the
resulting filter is guaranteed to be Stable. When two LSF
values are close to each other, a spectral peak is likely to occur
between them which is useful for tracking formants and
spectral peaks. In the training mode the source and target
speech is normalized to some predetermined amplitude range
and the pitch information is extracted to produce a residual,
residual contains vocal tract information which is modeled by
LPC. Applying the LPC analysis filter to the residual will
result in the vocal tract information being removed leaving
lung excitation signal.
LPC produces the results in an unstable synthesis filter. So we
convert LPC parameters into LSP. We have mapped the pitch
residual and LSP parameters of source speaker on to target
speaker using RBF neural network with spread factor of 0.01
and error threshold of 0.00001 respectively. In the
transformation phase the LSP and Pitch residual of test speech
samples are projected as an input to the trained MLP model,
the transformed LSP and pitch residuals are obtained. To
resynthesize the speech signal, LSP parameters are
reconverted into LPC, transformed speech is reconstructed
using LPC synthesis, as a target speech [3].
3.3 MLP QNN Based Transformation
Quantum Neural Networks are more efficient than classical
neural network for classification tasks, In that time required
for training is much less for QNN. The repetition is not
required by QNN; since each component networks learns only
one pattern causing the training set to be learned more quickly.
A QNN is shown in Figure 2. It has
n inputs in the form of qubits as
where
The output of the neuron is again a qubit

__________________________________________________________________________________________
Where is the output of the neuron for the kth
input set
is an operator that can be implemented by a network of
quantum gates. It has to satisfy the unitarity condition that
adj( ). = I (2)
is the weight matrix corresponding to the mth
input. It is of
the form
(3)
is the mth
input qubit and is of the form
(4)
The basic network considered in this paper is a generalized
MLP (GMLP) network, which consists of an input layer, an
output layer, a number of hidden nodes, and associated
interconnections. Let X and Y be the input and output vectors,
respectively. The GMLP network with m inputs, nh hidden
nodes, and n outputs is characterized by the equations
xi = Xi , 1 ≤ i ≤ m
, m < i ≤ m + nh + n
Yi=xi+m+nh, 1≤i≤n (5)
Fig 1: LSF based VC: Training and Testing model
where wi j is the weight of the connection from node j to node i
and f is the sigmoid function
(6)
The output unit performs weighted sum of hidden units. The
MLP Quantum neural network is trained to map LSP as a
vocal tract parameters and pitch residual (F0) as glottal
parameters between source and target speaker. Winner-takes-
all classification rule is used to adjust the weights of this
neural network so as to minimize the mean squared error
(MSE) between the desired and the actual output values. For
training, the LSP of the source speaker voice are used as input
and LSP of target speaker voice are used as a output of the
network. The weights of the hidden and output layers of the
networks are adjusted in such a way that the source speaker’s
patterns are converted into the target speaker’s patterns. The
training step is a supervised learning procedure [1].
Fig 2: QNN

__________________________________________________________________________________________
4. SIMULATION AND RESULT
Our algorithm is able to perform the mapping of the LSP, pitch
residual, of source to that of target speaker using MLP with
high accuracy. The sample results of our MLP based speech
conversion algorithm for inter gender speaker are shown in the
figures3. In figure 3, the left column represents the speech
signal waveform of source, target and transformed in order.
The right column in the figure 3 displays the Spectrogram of
source, target speaker and transformed speech in order from
top to bottom.
Fig 3: a) Source, b) Target and c) Transformed speech of
same sentence waveform of a male-to-female speaker of MLP
based speech transformation
The spectrograph is a two dimensional time and frequency
graphical plot of the energy present in the signal. Figure3,
display the waveforms and spectrograms for male to female
voice conversion. From the visual perception of the figure it is
clearly seen that the spectrogram of the target speaker is
similar to that of the transformed speech. The performance of
the QNN based transformation is better than RBF based
transformation. The proposed algorithm is able to transform
source speech to target speech successfully.
5 OBJECTIVE AND SUBJECTIVE
PERFORMANCE
In this section we evaluate our algorithm based on Objective
and subjective parameters. We use objective based evaluation
parameters (i) mean square error (MSE).
5.1 MSE based Objective Evaluation
In this section we provide the objective evaluation for MLP
based VC systems to measure the differences between the
target and transformed speech signals. Since many perceived
sound differences can be interpreted in terms of differences of
spectral features and mean squared error (MSE) are
considered to be a reasonable metrics; for both mathematically
and subjectively analysis. The MSE between target audio
vector c and transformed audio vector b is calculated as per
below equation on a sample by sample basis.
(7)
The average square difference between two vectors is used to
evaluate the objective performance of mapping algorithms.
5.2 Subjective Evaluation
Evaluation of the synthesized speech of desired speaker using
MLP and RBF based VC systems by standard subjective test
can be attributed due to inefficiency of objective evaluation
techniques in defining good objective distance measures
which are perceptually meaningful. In this paper subjective
evaluation tests have been discussed namely i) Mean Opinion
Score.
5.2.1 Mean Opinion Score
To evaluate the overall accuracy of the conversion, a listening
test is carried out for evaluating the similarity between the
converted voice and the target voice to find the performance of
MLP based transformation. Listeners give scores between 1
and 5 for measuring the similarity between the output of the
two VC systems and the target speaker’s natural utterances.
The results of this MOS similarity tests are provided in figure
6, which indicates that the MLP based voice conversion
systems completely characterized the characteristics of the
source speaker. The MLP based voice conversion has slightly
more resemblance to target speech as compared to the GMM
based VC system.
Fig 4: MOS Test
6. CONCLUSIONS
In this paper, we have proposed a novel technique using LSP
as spectral features and pitch residual as glottal features. The
MLP based and mapping functions are developed to map the
acoustical cues of the source speaker according to that of the
target speaker. The inter gender and intra gender VC is
performed using proposed algorithm. We have experimentally

__________________________________________________________________________________________
verified that, the fundamental frequency range of female
speech is higher than male speech and vice versa. The
comparative performance of RBF and MLP based models are
studied using objective and subjective measures .The
evaluation results indicate that MLP can be used as an
alternative to the RBF based transformation model. The
subjective evaluation convinced that the quality and
naturalness of the transformed speech can be achieved with
the proposed algorithm.
REFERENCES
[1] Tzyy-Chyang Lu, Gwo-Ruey Yu,and Jyh-Ching Juang,
“Quantum-Based Algorithm for Optimizing Artificial
Neural Networks”IEEE Transactions on Neural
Networks and Learning Systems, Vol. 24, No. 8,
August 2013.
[2] Tapas Kumar Patra , Biplab Patra ,
PuspanjaliMohapatra,”Text to Speech Conversion with
Phonematic Concatenation”, International Journal of
Electronics Communication and Computer Technology
(IJECCT) Volume 2 (September 2012).
[3] J.H.Nirmal, Suparva Patnaik , and Mukesh A.Zaveri,
“Line Spectral Pairs Based Voice Conversion using
Radial Basis Function”, ACEEE Int. J. on Signal &
Image Processing, Vol. 4, No. 2, May 2013.
[4] Chen Zhi ,Zhang Ling-hua“Voice Conversion Based on
Genetic Algorithms”,Telecommunication 2011.
[5] Daniel Erro, Asunción Moreno ,”Weighted Frequency
Warping for Voice Conversion,Speech”,
Communication2008.
[6] Danwen Peng,Xiongwei Zhang, ,Jian Sun,” Voice
Conversion Based on GMM and ArtificialNeural
Network” 2010 IEEE.
[7] Narpendyah W. Ariwardhani1, Yurie Iribe1, Kouichi
Katsurada1 and Tsuneo Nitta1, 2 ,“Voice Conversion
For Arbitrary Speakers Using Articulatory-Movement
To Vocal-Tract Parameter Mapping” ,2013 IEEE
International Workshop On Machine Learning For
Signal Processing, Sept. 22–25, 2013.
[8] R. Xuemei and L. Xiaohua, “Identification of extended
Hammersteinsystems using dynamic self-optimizing
neural networks,” IEEETrans.Neural Netw., vol. 22,
no. 8, pp. 1169–1179, Aug. 2011.
[9] Srinivas Desai, E.Veera Raghavendra,
B.Yegnanarayana, Alan W Black, Kishore
Prahallad,“Voice Conversion using Artificial
NeuralNetworks”in IEEE Int. Conf. on Acoustics,
Speech and Signal Processing (ICASSP), Taipei,
Taiwan, April 2009.
[10] Srinivas Desai, Alan W. Black, B. Yegnanarayana, and
Kishore Prahallad, “Spectral Mapping Using Artificial
Neura Networks For Voice Conversion” IEEE
TRANSACTIONS on Audio, Speech, And Language
Processing, VOL. 18, NO. 5, JULY 2010.
BIOGRAPHIES
A .DHIVYA BHARATHI is a student pursuing
PG Degree in Computer Science and
Engineering at knowledge institute of
technology, Salem. Research Interest includes
Artificial intelligence and Neural Networks
contact:dhivi.anand@gmail.com
T. KARTHIKEYAN obtained his Master degree
from Central Queensland University, Melbourne,
Australia in 2008. He is authored over 10
publications including Artificial Intelligence and
Wireless Sensor Networks and he has guided 3
international level projects. He is currently an Assistant
Professor in Knowledge Institute of Technology. His research
mainly focuses on Neural Networks and Cloud Computing.
Contact: tkcse@kiot.ac.in
PHEMALATHA is a student pursuing PG Degree
in Computer Science and Engineering at
knowledge institute of technology, Salem.
Research Interest includes Data Mining and
Artificial intelligence.
Contact:ghemalathaoct8@gmail.com
JOY KINSHY.P is a student pursuing PG Degree
in Computer Science and Engineering at
knowledge institute of technology, Salem.
Research Interest includes Mobile Computing and
Artificial intelligence.
Contact:cjoykinshy@gmail.com

Identification of frequency domain using quantum based optimization neural networks

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Identification of frequency domain using quantum based optimization neural networks (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

Identification of frequency domain using quantum based optimization neural networks