基于神经网络的语音识别研究

Research on Speech Recognition Based on Neural Networks

由于具有良好的抽象分类特性,神经网络现已应用于语音识别系统的研究和开发,并成为解决识别相关问题的有效工具。为解决一般语音识别系统准确率较低的问题,本文分别给出了由循环神经网络(RNN)和多层感知器(MLP)组成识别模块的两种语音识别系统,并对二者识别的准确性进行了比较。介绍了特征提取模块的主要工作步骤并讨论了组成识别模块的上述两种神经网络结构。其中,特征提取模块利用线性预测编码(LPC)倒谱编码器,把输入语音翻译成LPC倒谱空间中的曲线;而识别模块完成对某个特征空间曲线之间的联系和单词的识别。实验结果表明,MLP方法准确率高于RNN方法,而RNN方法准确率可达85%。

Because of good characteristics of the abstract classification, neural networks have become an effective tool for resolving issues related to recognition, and have been applied to the research and development of speech recognition systems. A speech recognizer system comprises of two blocks, Feature Extractor and Recognizer. For increasing the recognition accuracy, this paper proposes two types of speech recognition system whose recognition block uses the recurrent neural network（RNN） and multi layer pereeptron（MLP） respectively. Furthermore, the main work steps of Feature Extractor（FE） block is introduced and the structure of two types of neural networks mentioned above is discussed. Using a standard LPC Cepstrum, the FE translates the input speech into a trajectory in the LPC Cepstrum feature space. The recognizer block discovers the relationships between the trajectories and recognizes the word. The results show that the MLP＇s recognition accuracies were better than the RNN＇s, while the RNN＇s recognition accuracies achieved 85%.