计算机工程
計算機工程
계산궤공정
COMPUTER ENGINEERING
2015年
5期
213-218
,共6页
赵彩光%张树群%雷兆宜
趙綵光%張樹群%雷兆宜
조채광%장수군%뢰조의
对比散度%高斯伯努利受限玻尔兹曼机%受限玻尔兹曼机%指数平均数指标%并行回火%语音识别%深度神经网络
對比散度%高斯伯努利受限玻爾玆曼機%受限玻爾玆曼機%指數平均數指標%併行迴火%語音識彆%深度神經網絡
대비산도%고사백노리수한파이자만궤%수한파이자만궤%지수평균수지표%병행회화%어음식별%심도신경망락
Contrastive Divergence(CD)%Gaussian-Bernoulli Restricted Boltzmann Machine(GRBM)%Restricted Boltzmann Machine(RBM)%Exponential Moving Average(EMA)%Parallel Tempering(PT)%speech recognition%Deep Neural Network(DNN)
对比散度作为训练受限波尔兹曼机模型的主流技术之一，在实验训练中具有较好的测试效果。通过结合指数平均数指标算法和并行回火的思想，提出一种改进对比散度的训练算法，包括模型参数的更新和样本数据的采样，并将改进后的训练算法应用于高斯伯努利受限玻尔兹曼机( GRBM)中训练语音识别模型参数。在TI-Digits数字语音训练和数字测试数据库上的实验结果表明，采用改进的对比散度训练的GRBM明显优于传统的模型训练算法，语音识别率能够达到80%左右，最高提升7%左右，而且应用改进算法训练的其他GRBM对比模型的语音识别率也都有所提高，具有较好的识别性能。
對比散度作為訓練受限波爾玆曼機模型的主流技術之一，在實驗訓練中具有較好的測試效果。通過結閤指數平均數指標算法和併行迴火的思想，提齣一種改進對比散度的訓練算法，包括模型參數的更新和樣本數據的採樣，併將改進後的訓練算法應用于高斯伯努利受限玻爾玆曼機( GRBM)中訓練語音識彆模型參數。在TI-Digits數字語音訓練和數字測試數據庫上的實驗結果錶明，採用改進的對比散度訓練的GRBM明顯優于傳統的模型訓練算法，語音識彆率能夠達到80%左右，最高提升7%左右，而且應用改進算法訓練的其他GRBM對比模型的語音識彆率也都有所提高，具有較好的識彆性能。
대비산도작위훈련수한파이자만궤모형적주류기술지일，재실험훈련중구유교호적측시효과。통과결합지수평균수지표산법화병행회화적사상，제출일충개진대비산도적훈련산법，포괄모형삼수적경신화양본수거적채양，병장개진후적훈련산법응용우고사백노리수한파이자만궤( GRBM)중훈련어음식별모형삼수。재TI-Digits수자어음훈련화수자측시수거고상적실험결과표명，채용개진적대비산도훈련적GRBM명현우우전통적모형훈련산법，어음식별솔능구체도80%좌우，최고제승7%좌우，이차응용개진산법훈련적기타GRBM대비모형적어음식별솔야도유소제고，구유교호적식별성능。
Contrastive divergence has a good result for training restricted Boltzmann machine model as one of the mainstream training algorithm in the experiments. An improved contrastive divergence based on Exponential Moving Average ( EMA ) is proposed by combining with the exponential moving average learning algorithm and Parallel Tempering( PT) ,which includes updating the model parameters and samples. The improved algorithm is applied to train speech recognition model parameters in Gaussian-Bernoulli Restricted Boltzmann Machine ( GRBM ) , and experimental results of digit speech recognition on the core test of TI-Digits show that the proposed algorithm works better than traditional training algorithms in GRBM,the accuracy can be as high as 80. 53% and increase by about 7%. Recognition accuracy of some other GRBM models also increase apparently based on the proposed algorithm. And its performance keeps well.
