Tagging Tone for Mandarin Pinyin Based on Sequence Labelling

Abstract

Generally, fundamental frequency (F0) is applied as the clue for tone recognition in normal speech. Tone recognition in whisper speech without F0 could be based on the temporal and spectral cues. However, the Mandarin Electro-Laryngeal (EL) speech with fixed F0 has no tone information. Therefore, the tone recognition of Mandarin EL speech is so difficult. And the researches about tone recognition for Mandarin EL speech is insufficient. In this paper, a new method labelling the tone for pinyin is proposed based on the context information to identify the tone of Mandarin speech without tone information. The experiment result shows that the precision, recall and the F value are all above 97% based on the test dataset. The amount of semantic information influences the performance of proposed method. If the amount of semantic information is little, the accuracy of tone labelling would be poor. The result shows that the proposed method has a good precision and robustness. The method can label the tone for pinyin without any tone information only based on the context information. The proposed method can label tones for tonal language, such as the Mandarin speech.