Syllable Segmentation of Pumi Speech with GABPNN Based on Fractal Dimension and DWPTMFCC

In this paper, a method is put forward to implement the automatic syllable segmentation of continuous Pumi speech with GABPNN, based on the two features of DWPTMFCC and fractal dimension. Concretely, speech signal is at first denoised by wavelet packet transform. Afterwards, for speech and non-speech frames, fractal dimension of each frame is extracted and labeled respectively. The same is also done for DWPTMFCC by wavelet packet transform. As a feature vector, a 17 dimensional vector from 17 DWPTMFCCs, together with fractal dimension is used to train and test GABPNN. Finally, according to DWPTMFCC, the automatic syllable segmentation of continuous Pumi speech is implemented, combining with dynamic threshold of fractal dimension. The experimental results show that the average accuracy of the auto segmentation arrives at 95.16%, and that our method is effective, and better than conventional ones with regard to continuous Pumi speech.