为建立更高效的电力智能交互式平台,需要更准确地识别参与人的交互意图。针对目前使用的基于简单关键词匹配的意图识别方法准确率较低的问题,提出一种基于压缩时延神经网络(compressed Time Delay Neural Network,cTDNN)与卷积神经网络(Convolutional Neural Network,CNN)的语音关键点定位算法。算法通过引入延时单元有效降低传统方法的时间复杂度,提升意图识别的实时性。同时,通过引入卷积神经网络,学习语音中丰富的上下文相关性,提升关键点定位的准确性。在真实场景中采集的交互数据集上的实验表明,算法可有效提升电力智能交互式场景中对参与人意图的识别能力。
英文摘要:
For building an intelligent power interaction platform with higher efficiency, it is quite necessary to recognize the intention of the participant with higher accuracy. To overcome the shortcomings of low accuracy in the present keyword matching based intention recognition methods, a compressed time delay neural network (cTDNN) and convolutional neural network (CNN) is proposed. Time delay units are introduced to reduce the high computational complexity and improve the performance of intention recognition. By exploiting convolutional neural network, contextual semantical information in voice signal can be learned, and accuracy of keyword spotting can be increased. Results of experiments on dialog dataset collected from real context indicate that the proposed method efficiently improve the ability of intention recognition in intelligent interaction platform for power.
