基于深度神经网络的电力调度语音识别研究及应用

Research and Application of Power Dispatching Speech Recognition Based on Deep Neural Network

为进一步提高智能电网调度语音识别的准确率,本文将深度学习技术引入了电力调度语音识别领域,提出了基于深度神经网络(deep neural network,DNN)的电力调度语音识别技术。针对电力调度专用术语以及某区域电网习惯调度用语,建立了电力调度语音识别基础语料库。通过应用深度神经网络-隐马尔可夫模型(deep neural network-hidden markov model,DNN-HMM)进行声学模型训练,并对电力调度语音材料制作,语音模型训练过程、端点检测、与D5000系统交互以及语音转文字的整套流程进行了论述。实践结果表明,采用DNN-HMM的电力调度语音识别性能要显著优于传统语音识别框架,即高斯混合-隐马尔可夫模型(gaussian mixture model-hidden markov model,GMM-HMM),采用所提方法进行电力调度语音识别准确率达94.63%。基于所提方法开发的电力调度语音识别系统在某区域电网调控中心的应用实例表明了所提方法的可行性与优良性。

In order to further improve the accuracy of smart grid power dispatching speech recognition,this paper introduces deep learning technology into the field of power dispatch speech recognition,and proposes a power dispatching speech recognition technology based on deep neural network(DNN).A basic corpus of power dispatching speech recognition is established for the terminology and idiom of power dispatching in a certain regional power grid.This paper also uses the deep neural network-hidden markov model(DNN-HMM)for acoustic model training,and discusses the production of power dispatching voice materials,training process of voice model,endpoint detection,interaction with D5000 system and the entire process of voice-to-text.The practice results show that the speech recognition performance of power dispatching using DNN-HMM is significantly better than that of the traditional speech recognition framework,namely Gaussian mixture model-hidden markov model(GMM-HMM);the speech recognition accuracy of the proposed method is 94.63%,and an application example of a power dispatching speech recognition system based on the proposed method in a regional power grid control center shows the feasibility and superiority of the proposed method.