ISCA Archive Interspeech 2016
ISCA Archive Interspeech 2016

Tone Classification in Mandarin Chinese Using Convolutional Neural Networks

Charles Chen, Razvan Bunescu, Li Xu, Chang Liu

In tone languages, different tone patterns of the same syllable may convey different meanings. Tone perception is important for sentence recognition in noise conditions, especially for children with cochlear implants (CI). We propose a method that fully automates tone classification of syllables in Mandarin Chinese. Our model takes as input the raw tone data and uses convolutional neural networks to classify syllables into one of the four tones in Mandarin. When evaluated on syllables recorded from normal-hearing children, our method achieves substantially higher accuracy compared with previous tone classification techniques based on manually edited F0. The new approach is also more efficient, as it does not require manual checking of F0. The new tone classification system could have significant clinical applications in the speech evaluation of the hearing impaired population.