| 基于RBM的语音特征提取方法研究 |
| |
| 引用本文： | 赵从健,雷菊阳,李明明.基于RBM的语音特征提取方法研究[J].教育技术导刊,2020,19(2):114-117. |
| |
| 作者姓名： | 赵从健 雷菊阳 李明明 |
| |
| 作者单位： | 上海工程技术大学 机械与汽车工程学院，上海 201620 |
| |
| 摘 要： | 针对传统语音识别在多目标情况下识别率较低的问题，从特征参数提取角度，提出一种基于受限玻尔茨曼机（RBM）的特征提取方法。依据不同个体语音信号之间的特征差异提取特征参数，通过梯度上升算法调整网络参数以拟合给定训练样本，通过对比散度算法降低采样达标所需状态转移次数以提高算法效率，再利用重构误差曲线评价受限玻尔茨曼机对训练样本的似然度。实验表明，当隐含层节点个数为30时，参数提取的重构误差低于20%。此时使用改进的BP网络训练，与传统算法相比，综合识别率提高到86.9%，对提升多目标语音识别率具有重要意义。
|
| 关 键 词： | 语音识别 受限玻尔茨曼机 特征提取 梯度上升 对比散度 |
| 收稿时间： | 2019-04-04 |
Research on Speech Feature Extraction Method Based on Restricted Boltzmann Machine |
| |
| ZHAO Cong-jian,LEI Ju-yang,LI Ming-ming.Research on Speech Feature Extraction Method Based on Restricted Boltzmann Machine[J].Introduction of Educational Technology,2020,19(2):114-117. |
| |
| Authors: | ZHAO Cong-jian LEI Ju-yang LI Ming-ming |
| |
| Institution: | College of Mechanical and Automotive Engineering，Shanghai University of Engineering Science，Shanghai 201620，China |
| |
| Abstract: | In order to solve the problem that the rate of traditional speech recognition was low in the case of multiple targets， from the point view of feature parameter extraction， a feature extraction method based on restricted Boltzmann machine was proposed. It extracted the characteristic parameters mainly based on the characteristic differences between different individual speech signals， adjusted the parameters of the network by the gradient rise algorithm to fit the given training sample， reduced the number of state transitions required for sampling to reach the standard by the contrast divergence algorithm to improve the efficiency of the algorithm， used the reconstruction error curve to evaluate the likelihood of the restricted Boltzmann machine to the training samples. Experiments showed that when the number of hidden layer nodes is 30， the reconstruction error is less than 20%. Compared with traditional algorithm， the comprehensive recognition rate obtained form the improved BP network training was raised to 86.9%， which was of great significance for improving the speech recognition rate of multiple targets. |
| |
| Keywords: | speech recognition restricted Boltzmann machine feature extraction gradient rising contrast divergence |
|
| 点击此处可从《教育技术导刊》浏览原始摘要信息 |
| 点击此处可从《教育技术导刊》下载免费的PDF全文 |
