DETECTING NONLINEARITIES IN SPEECH SOUNDS USING THE BICOHERENCE
JWA FACKRELL, Stephen McLaughlin · 2024
A signal property called quadratic phase coupling (QPC) is the signature ofa quadratically nonlinear signal production mechanism.and tools from the eld ofhigher-order statistics (HOS) can be used to detect this property.This paper describes an investigation into the use of a HOS tool called the hicoherence for detecting QPC in continuant speech signals.A detector for QPC is dscribed, and applied to a number of speech sounds from a specially-recorded database.The results are consistent with the hypotheses that unvoiced fricativa areGaussian, and that voiced continuants (vowels, voiced fricatives and nasals) do not exhibit QPC.Therefore it seems that quadratic models of speech signals are unlikely to offer improvements over existing, linear.models. . INTRODUCTIONMost speech synthesizers are based on the linear model of speech production.and this model, although mathematically simple.is capable of dacribing a great many of the properties of human speech.Indeed much of the speech generated by linear synthesizers is intelligible, which for many applications is all that is required.However, synthesized speech does not sound very natural, and this problem has led several researchers to investigate the possibilities of using nonlinear models for speed: production.Kubin [l] provida a comprehensive review of these.One of the problems encountered in attempts to develop nonlinear models for speech production is that of choosing what type of nonlinear model is appropriate.A viable rst step in such work is to try to identify the signatures of certain type of nonlinearity in human speech signals.If thse signatures are found, then they provide clues to what type of nonlinear speech production model is appropriate.This paper is concerned with the identi cation of the signature quadratic nonlinean'ty, which is just one particular type of nonlinearity.Signals which are produced by modelscontaining quadratic nonlinearities exhibit a property called quadratic phase coupling (QPC).The idea is that if QPC is detected in real speech.then models including quadratic nonlinearities may be good ones for speech signals.This QPC phenomenon can be detected using measures fromthe eld of higher-order statistics (HOS).HOS techniques can provide information about signal phase, nonlinearities and deviations 'During the period in which this work was carried out.Justin Far:de was supported by a CASE studentahip from EPSRC in collaboration with BT Laboratoriu.Martimhun Heath.Steve McLaughlin is