CAJ | 학술논문

研究将深度神经网络有效地应用到维吾尔语大词汇量连续语音识别声学建模中的两种方法：深度神经网络与隐马尔可夫模型组成混合架构模型(Deep neural network hidden Markov model,DNN-HMM),代替高斯混合模型进行状态输出概率的计算；深度神经网络作为前端的声学特征提取器提取瓶颈特征(Bottleneck features,BN),为传统的 GMM-HMM(Gaussian mixture model-HMM)声学建模架构提供更有效的声学特征(BN-GMM-HMM)。实验结果表明,DNN-HMM 模型和 BN-GMM-HMM模型比 GMM-HMM 基线模型词错误率分别降低了8.84%和5.86%,两种方法都取得了较大的性能提升。
연구장심도신경망락유효지응용도유오이어대사회량련속어음식별성학건모중적량충방법：심도신경망락여은마이가부모형조성혼합가구모형(Deep neural network hidden Markov model,DNN-HMM),대체고사혼합모형진행상태수출개솔적계산；심도신경망락작위전단적성학특정제취기제취병경특정(Bottleneck features,BN),위전통적 GMM-HMM(Gaussian mixture model-HMM)성학건모가구제공경유효적성학특정(BN-GMM-HMM)。실험결과표명,DNN-HMM 모형화 BN-GMM-HMM모형비 GMM-HMM 기선모형사착오솔분별강저료8.84%화5.86%,량충방법도취득료교대적성능제승。
Two methods are proposed by employing deep neural network for Uyghur large vocabulary con-tinuous speech recognition :Hybrid architecture models are established with deep neural network (DNN) and hidden Markov model (HMM)for replacing Gaussian mixture model (GMM)in GMM-HMM to compute the state emission probabilities;DNN is facilitated as a front-end acoustic feature extractor to extract bottleneck feature(BN)to provide more effective acoustic features for the traditional GMM-HMM modeling framework (BN-GMM-HMM).The experimental results show that DNN-HMM and BN-GMM-HMM reduce word error rate(WER)by 8.84% and 5.86% compared with the GMM-HMM base-line system,which demonstrates that the two methods accomplish significant performance improve-ments.