基于深度学习的语音识别模型及其在智能家居中的应用
|
摘要
为了满足人们对智能家居设备控制便捷化的需求,提出了一种基于降噪自动编码器的深度学习语音识别模型,经过语音识别模型解析出短语控制指令,以实现家居设备控制。该语音识别模型主要包含两个部分:首先进行无监督学习预训练,预训练前随机将一些网络节点置为0,人工模拟噪声数据,然后采用限制玻尔兹曼机权重矩阵依次训练每一个隐含层,通过比较输入数据与输出数据的偏差修改权重,优化参数;然后进行有监督微调,把训练好的参数作为整个网络的初始值,采用误差反向传播算法对整个网络模型调参。实验结果表明:该语音识别模型与深度信念网络对比,在语音识别率和对噪声的鲁棒性都有明显提高。将该语音识别模型和智能家居系统相结合,从普通短语中判断出家居控制指令,实现人机交互非接触式、便捷式控制,从而使系统更加智能化。
In order to meet the needs of people to control smart home equipment conveniently, a deep learning speech recognition model based on denoising autoencoder was proposed. Through the speech recognition model, the phrase control instruction was parsed to achieve the purpose of home equipment control. The speech recognition model mainly consists of two parts. The first part is unsupervised learning pre-training. Before the unsupervised pre-training, some network nodes were randomly set to 0; the noise data were artificially simulated; then each hidden layer was trained sequentially by using the Boltzmann machine weight matrix. The weight was modified and the parameters were optimized through comparing the deviation between input data and output data. Then, supervised fine adjustment was conducted. The well-trained parameters served as the initial values of the whole network, and error back propagation algorithm was adopted to adjust parameters of the whole network model. The experimental results showed that speech recognition rate and noise robustness of the speech recognition model improve significantly, compared with deep belief network. The speech recognition model could be combined with smart home system to judge home control command from the common phrase and achieve human-computer interaction non-contact and convenient control so that the system is more intelligent.
In order to meet the needs of people to control smart home equipment conveniently, a deep learning speech recognition model based on denoising autoencoder was proposed. Through the speech recognition model, the phrase control instruction was parsed to achieve the purpose of home equipment control. The speech recognition model mainly consists of two parts. The first part is unsupervised learning pre-training. Before the unsupervised pre-training, some network nodes were randomly set to 0; the noise data were artificially simulated; then each hidden layer was trained sequentially by using the Boltzmann machine weight matrix. The weight was modified and the parameters were optimized through comparing the deviation between input data and output data. Then, supervised fine adjustment was conducted. The well-trained parameters served as the initial values of the whole network, and error back propagation algorithm was adopted to adjust parameters of the whole network model. The experimental results showed that speech recognition rate and noise robustness of the speech recognition model improve significantly, compared with deep belief network. The speech recognition model could be combined with smart home system to judge home control command from the common phrase and achieve human-computer interaction non-contact and convenient control so that the system is more intelligent.
引文
[1] 王爱芸.语音识别技术在智能家居中的应用[J].软件,2015(7):104-107.
[2] 闵梁.面向智能家居的语音识别技术研究与实现[D].哈尔滨:哈尔滨工业大学, 2013:25-49.
[3] 刘文强.语音识别技术在智能家居中的研究与应用[D].大连:大连海事大学,2013:15-62.
[4] 戴礼荣,张仕良,黄智颖.基于深度学习的语音识别技术现状与展望[J].数据采集与处理,2017,32(2):221-231.
[5] 居治华,刘罡,陈琦岚,等.基于反向卷积的 Bi-LSTM 语音识别[J]. 软件导刊, 2017(7): 27-30.
[6] 杨洋,汪毓铎.基于改进卷积神经网络算法的语音识别[J].应用声学,2018(6):940-946.
[7] 包晓安,常浩浩,徐海,等.基于LSTM的智能家居机器学习系统预测模型研究[J].浙江理工大学学报, 2018,39(2): 224-231.
[8] 包晓安,林辉,周建平,等.基于智能家居的一致性模型融合技术研究[J].浙江理工大学学报,2015, 33(1): 109-114.
[9] 童江松,李仁旺,钱小燕.基于ARM的智能家居红外控制系统设计[J].浙江理工大学学报, 2015, 33(1): 124-129.
[10] 侯一民,周慧琼,王政一.深度学习在语音识别中的研究进展综述[J].计算机应用研究,2017,34(8): 2241-2246.
[11] Li Q, Huang Y. An auditory-based feature extraction algorithm for robust speaker identification under mismatched conditions[J]. IEEE Transactions on Audio, Speech, and Language Processing, 2011, 19(6): 1791-1801.
[12] 王山海,景新幸,杨海燕.基于深度学习神经网络的孤立词语音识别的研究[J].计算机应用研究,2015, 32(8): 2289-2291.
[13] 蒋泰,张林军.语音识别自适应算法在智能家居中的应用[J].计算机系统应用, 2017, 26(3):150-155.
[14] Zhang S X, Liu C, Yao K, et al. Deep neural support vector machines for speech recognition[C]// IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2015: 4275-4279.
[15] Maas A L, Qi P, Xie Z, et al. Building DNN acoustic models for large vocabulary speech recognition[J]. Computer Speech & Language, 2017, 41:195-213.
[2] 闵梁.面向智能家居的语音识别技术研究与实现[D].哈尔滨:哈尔滨工业大学, 2013:25-49.
[3] 刘文强.语音识别技术在智能家居中的研究与应用[D].大连:大连海事大学,2013:15-62.
[4] 戴礼荣,张仕良,黄智颖.基于深度学习的语音识别技术现状与展望[J].数据采集与处理,2017,32(2):221-231.
[5] 居治华,刘罡,陈琦岚,等.基于反向卷积的 Bi-LSTM 语音识别[J]. 软件导刊, 2017(7): 27-30.
[6] 杨洋,汪毓铎.基于改进卷积神经网络算法的语音识别[J].应用声学,2018(6):940-946.
[7] 包晓安,常浩浩,徐海,等.基于LSTM的智能家居机器学习系统预测模型研究[J].浙江理工大学学报, 2018,39(2): 224-231.
[8] 包晓安,林辉,周建平,等.基于智能家居的一致性模型融合技术研究[J].浙江理工大学学报,2015, 33(1): 109-114.
[9] 童江松,李仁旺,钱小燕.基于ARM的智能家居红外控制系统设计[J].浙江理工大学学报, 2015, 33(1): 124-129.
[10] 侯一民,周慧琼,王政一.深度学习在语音识别中的研究进展综述[J].计算机应用研究,2017,34(8): 2241-2246.
[11] Li Q, Huang Y. An auditory-based feature extraction algorithm for robust speaker identification under mismatched conditions[J]. IEEE Transactions on Audio, Speech, and Language Processing, 2011, 19(6): 1791-1801.
[12] 王山海,景新幸,杨海燕.基于深度学习神经网络的孤立词语音识别的研究[J].计算机应用研究,2015, 32(8): 2289-2291.
[13] 蒋泰,张林军.语音识别自适应算法在智能家居中的应用[J].计算机系统应用, 2017, 26(3):150-155.
[14] Zhang S X, Liu C, Yao K, et al. Deep neural support vector machines for speech recognition[C]// IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2015: 4275-4279.
[15] Maas A L, Qi P, Xie Z, et al. Building DNN acoustic models for large vocabulary speech recognition[J]. Computer Speech & Language, 2017, 41:195-213.