Detecting Tone Errors In Continuous Mandarin Speech

Yan-Bin Zhang,Min Chu,Chao Huang,Man-Gui Liang
DOI: https://doi.org/10.1109/ICASSP.2008.4518797
2008-01-01
ICASSP
Abstract:This paper proposes a new approach for detecting tone errors in continuous Mandarin speech. In the training phase, tone variations are modeled with context-depended MSD-HMM which considers six contextual factors instead of two in traditional triphone HMM. 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%.
What problem does this paper attempt to address?