Speech Segmentation of Chinese Dialects and Accented Mandarin by Initials and Finals Templates

Speech segmentation is an important key step and fundamental work in language researches and intelligent applications. When performing speech segmentation on low resource languages, the method based on unit recognition is difficult to construct the acoustic model and language model due to lack of training data, while the method based on boundary detection has poor generalization ability and stability due to setting parameters. The method based on template matching seems to be able to overcome these problems by directly identifying language units in speech data through samples, completely abandoning the model and training process. Unlike template-based speech recognition, we only use templates to detect possible speech units and determine their boundaries, rather than recognizing specific words. Therefore, the smaller and fewer speech units can be focus on and built templates, which is very suitable for low resource speech, such as Chinese dialects and accented Mandarin. Based on the fact that both Chinese Mandarin and Chinese dialects are monosyllabic languages, this paper establishes the reference templates for initials and finals. By using template matching and rule filtering, the boundaries of initials and finals are determined and finally the speech data are divided into multiple syllables. The proposed algorithm has been validated on the speech data of accented Mandarin and Chinese dialects, and the comparison with other algorithms have also been performed on these data. The average segmentation accuracy has achieved over 90% on the tolerance threshold of 20ms.