Detecting Tone Errors In Continuous Mandarin Speech

In recent years,much progress has been made in the area of computer-assisted language learning(CALL)system,in which pronunciation evaluation plays an important role.Yet,only a few works have been done in evaluating Mandarin pronunciation and most of them are on segmental goodness.Since Mandarin is a tonal language,it is very important to pronounce tone precisely in live communication,and therefore,detecting tone errors is crucial for a Mandarin CALL system.In this paper, we proposed to detect tonal errors by measuring the Kullback-Leibler Divergence(KLD) between the expected tone model and the most representative tone model(the tone model that matched real speech the most).And we proposed to model tone variations by Context-Dependent Tone Model(CDTM).In a previous work,Zhang et.al.used log-posterior probability as a measure of goodness of tone pronunciation.In a monosyllabic corpus,they got about 90%accuracy allowing 4%false acceptance rates.Si Wei used a similar approach but with F0 after CDF-matching normalization as the feature to detect tone errors.The Cross-Correlation between human experts and automatic tone error detection system is close to 0.79.Both works modeled tones with triphone Hidden Markov Models(HMM)and achieved promising results on isolated syllables.In this paper,we focus on tone error detection in continuous speech and propose to model tone variations with context-dependent HMMs.For a continuous speech segment,a sequence,of expected CDTMs is derived from the corresponding script and a sequence of most representative CDTMs is generated by model selection against the speech.We propose to measure the goodness of tone pronunciation by the KLD between the expected model and representative model.In the evaluation phase,the goodness of tone pronunciation is measured by Kullback-Leibler Divergence(KLD)between the expected tone model and the most representative tone model.When the KLD between the two models is larger than a threshold,the tone is detected as a pronunciation error.In the ROC curve,we get the equal error rate at 2.6%.