Applying Deep Bidirectional Long Short-Term Memory To Mandarin Tone Recognition

Automatic tone recognition is an important module of Computer-Aided Pronunciation Training (CAPT) system and it can provide many instructive information to guide the learner to practice their tone pronunciation in the process of Mandarin Chinese learning. However, many factors including the tone sandhi, the tonal articulation and other effects have many effects on tone recognition in continuous speech. To deal with these factors, we explored the Deep Bidirectional Long Short-Term Memory (BLSTM), which can utilize the context information from bi-direction in the time series, to Mandarin tone recognition. In addition, convolutional layers were adopted to reduce the spectral variation. The experimental results showed that the performance of proposed CNN-DBLSTM was the best and it achieved the tone error rate (TER) of 7.03% with a 15.1% relative error reduction from the DNN baseline system with TER of 8.27%. It demonstrated that our proposed model was more effective to handle the F0 variations than other models.