Speaker-independent automatic detection of pitch accent

This paper presents a novel approach to the automatic detection of pitch accent in spoken English. The approach that we propose is based on a time-delay recursive neural network (TDRNN), which takes into account contextual information in two ways: (1) a delayed version of prosodic and spectral features serve as inputs which represent an explicit trajectory along time; and (2) recursions from the output layer and some hidden layers provide the contextual labeling information that reflects characteristics of pitch accentuation in spoken English. We apply the TDRNN to pitch accent detection in two forms. In the normal TDRNN, all of the prosodic and spectral features are used as an entire set in a single TDRNN. In the distributed TDRNN, the network consists of several TDRNNs each treating each prosodic feature as a single input. In addition, we propose a feature called spectral balance-based cepstral coefficient (SBCC) to capture the spectral characteristic of pitch accentuation. We used the Boston Radio News Corpus (BRNC) to conduct experiments on the speakerindependent detection of pitch accent. The experimental results showed that the automatic labels of pitch accent exhibited an average of 83.6% agreement with the hand labels.

本文提出了一种自动检测英语口语音高重音的新方法。我们提出的方法基于时延递归神经网络（TDRNN），它通过两种方式考虑语境信息：（1）韵律和频谱特征的延迟版本作为输入，代表了沿时间的明确轨迹；（2）来自输出层和一些隐藏层的递归提供了反映英语口语音高重音特征的语境标注信息。我们将TDRNN以两种形式应用于音高重音检测。在常规的TDRNN中，所有的韵律和频谱特征在单个TDRNN中作为一个整体集合使用。在分布式TDRNN中，网络由几个TDRNN组成，每个TDRNN将每个韵律特征作为单个输入。此外，我们提出了一种称为基于频谱平衡的倒谱系数（SBCC）的特征来捕捉音高重音的频谱特征。我们使用波士顿广播新闻语料库（BRNC）对音高重音的与说话人无关的检测进行了实验。实验结果表明，音高重音的自动标注与人工标注的平均一致性为83.6%。