使用MLP與韻律模型之聲調辨認

Abstract

在本論文中，基本辨識系統上，對單一音節的辨認運用前後音節的特徵參數，並對於音高輪廓及能量區段化的方式，利用MLP辨認器進行聲調辨認，實驗於單一語者及非特定語者語料庫，辨認率分別為87.74%及83.27%；擴展特徵參數抽取方式至tone pair上，同樣利用MLP辨認器，加上利用Viterbi search對於MLP辨認器進行修正，辨認率分別為88.15%及85.81%；此外，利用音節的基頻軌跡、音節間的pause duration及energy-deep level，訓練聲調模型、韻律模型、音節間的break type模型，利用Viterbi search做聲調辨認，對於單一語者語料庫，最高可得到辨識為71.89%。

In this thesis, the features of the preceding and the succeeding syllable are used to help tone recognition on MLP (multi-layer perceptron) tone recognizer. The features include means and slopes of three uniformly divided-pitch contour, duration of the syllable and energy. Recognition rate are 87.74% and 83.27% for single speaker and multi-speaker database. If using the features of tone pair on MLP tone recognizer, the recognition rate are 88.15% and 85.81% respectively. Furthermore, using the features of pitch contour, pause duration and energy-dip level construct prosody model, tone model and break type model. Then we use Viterbi search algorithm to recognize. A recognition rate of 71.89% is achieved.