Abstract: A novel phonological attribute detection method based on long-term information is presented. This method is comprised of high-level and low-level Time-delayed Neural Networks(TDNN). The low-level TDNN carries out phonological attribute detection on the basis of short-term features, and the high-level TDNN is based on the low-level output and considering the long-term information, and fully taps the relation between speech signals in time. Experimental results show that, compared by the detection using short-term features, the introduction of phonological attribute based on long-term features improves detection rate with 3%. In addition, this paper puts the phonological attribute in phoneme recognition experiments, the results improveing 1.7% in Hidden Markov Model(HMM)-based speech recognition system.

Key words: phonological attribute, long-term features, hierarchical structure, Artificial Neural Network(ANN), Hidden Markov Model(HMM), phoneme classification

摘要： 提出一种基于长时性信息的音位属性检测方法，该方法通过高、低两层时间延迟神经网络(TDNN)进行实现，低层TDNN在短时特征上进行音位属性的检测，高层TDNN在低层检测结果的基础上，对更长时段上的信息进行融合。实验结果表明，引入长时性特征使得音位属性检测率提升约3%，将音位属性后验概率作为音素识别系统的观测特征，使用长时性特征的识别结果提升约1.7%。

关键词: 音位属性, 长时特征, 层级结构, 人工神经网络, 隐马尔可夫模型, 音素识别

CLC Number: 

• TP391.4