Accent Classication Using Support Vector Machine and Hidden Markov Model

zoomzurichAI and Robotics

Oct 16, 2013 (3 years and 7 months ago)


Accent Classication Using Support Vector
Machine and Hidden Markov Model
Hong Tang and Ali A.Ghorbani
Faculty of Computer Science
University of New Brunswick
Fredericton,NB,E3B 5A3,Canada
Abstract.Accent classication technologies directly in uence the per-
formance of speech recognition.Currently,two models are used for accent
detection namely:Hidden Markov Model (HMM) and Articial Neural
Networks (ANN).However,both models have some drawbacks of their
own.In this paper,we use Support Vector Machine (SVM) to detect dif-
ferent speakers'accents.To examine the performance of SVM,Hidden
Markov Model is used to classify the same problemset.Simulation results
show that SVMcan eectively classify dierent accents.Its performance
is found to be very similar to that of HMM.
1 Introduction
Accent is one of the most important characteristics of speakers.Recently,accent
detection became more focused and a number of researchers have published
works not only on the features of foreign accent but also on accent identication.
Levent and Hansen used Hidden Markov Model (HMM) codebooks based on the
acoustic features to identify three accents (American,Turkish,Chinese) aecting
English[?].Currently,extensive researches are being carried out to nd a suitable
method that can eectively detect speakers'accents.There are two main factors
{ acoustic features and the classication models.This paper proposes another
classication model-Support Vector Machine (SVM) to detect the accents.
2 Support Vector Machine
Besides HMM,ANN is the most popular model used to detect accents.Unfortu-
nately,ANNsuers fromnumber of limitations such as overtting,xed topology
and slow convergence.Statistical learning techniques based on risk minimization
such as Support Vector Machine (SVM) are found to be very powerful classica-
tion schemes.Compared with ANN,SVMhas several merits:(1) Structural Risk
Minimization techniques minimize a risk upper bound on the VC-dimension,(2)
among all hyperplanes separating the data,SVM can nd a unique hyperplane
that maximizes the margin of separation between the classes and (3) the power
of SVM lies in using kernel function to transform data from the low dimension
space to the high dimension space and construct a linear binary classier.
2 Hong Tang and Ali A.Ghorbani
In general,SVMis a binary classier.Recently,researchers have expanded the
basic SVM to the multi-class SVM.Such multi-class SVM has been successfully
applied to dierent kinds of classication problems.In our experiments,we use
Pairwise SVM and DAGSVM to classify three accents { Canadian,Chinese and
Indian accents.Pairwise Multi-class SVMis also called 1-to-rest algorithm.DAG
multi-class method is 1 to 1 algorithm,which uses a Directed Acyclic Graph
(DAG) to construct a binary tree.
3 Implementation
3.1 Speech Signal Database:The speech database consists of 60 male speak-
ers speech signals (20 Chinese speakers,20 Canadian speakers and 20 Indian
speakers).The choice of collecting speech data of one gender type (i.e.male
speakers) reduces the in uence of the dierent pitch frequency that exists be-
tween males and females.
3.2 Feature Extraction:Foreign accent is a pronunciation feature of non-
native speakers.Particular speech background groups generally exhibit some com-
mon acoustic features.We can identify dierent accent groups according to such
features.There are four features we used in our experiments.(1) Word-nal
Stop Closure Duration:the duration of the silence between the lax stop and
the full stop;(2) Word Duration:the time between the start and end of the
speech signal;(3) Intonation:the intonation depends on the syntax,semantics,
and phonemic structure of particular language;(4) F2-F3 contour:presents dif-
ferent tongue movements.The latter is the most powerful feature to distinguish
dierent accents.
3.3 Experiment Results And Conclusions:To examine the performance of
the SVM,we used HMM to detect the same accents.The test results are shown
in Table 1.From the results obtained,we found that:Pairwise SVM is not as
good as DAGSVM;DAGSVM has almost the same performance as HMM in
three accents database;the simulation results show that SVM and HMM have
almost the same convergence speed.
Table 1.Detection Rates for SVM and HMM
Pairwise SVM
References1.Levent Arslan,John H.L.Hansen Language Accent Classication in American
English Univeristy of Colorado Boulder,Speech Communication,Vol.18(4),pp.353-
367,July 1996.
2.Levent M.Arslan,John H.L.Hansen A study of Temporal Features and Frequency
Characteristic in American English Foreign Accent,Duck University,Robust Speech
Processing Laboratory.
3.John C.Platt,Nello Cristianini,John Shawe-Taylor Large Margin DAGs for Mul-
ticlass Classication