Research On Speaker Adaptation Method Based On DNN Acoustic Model

With the rapid development of neural networks and speech recognition,in-depth research on speaker adaptive technology in speech recognition has received more and more attention.The speaker adaptive method based on DNN acoustic model is researched.The acoustic model is adaptively trained according to the speaker's adaptive data,so as to improve the adaptive ability of the acoustic model to the speaker,so that the recognition system obtains the ratio.Adapt to higher recognition accuracy.At the same time,the channel interference in the adaptive process is studied to improve the robustness of the system.All in all,speaker adaptation has important research value.In the research of speaker adaptive method based on DNN acoustic model,this paper mainly studies the training and extraction of characterizing speaker identity(identity-vector,ivector).In order to reduce the influence of channel difference,the channel compensation method of i-vector feature is studied.Then,the adaptive training method of DNN acoustic model is studied.The details are as follows:Firstly,aiming at the over-fitting problem in adaptive data sparse training,this paper proposes a low-dimensional feature extraction technique based on Singular Value Decomposition(SVD),and gives the weight matrix in DNN network.The SVD decomposition formula and the corresponding analysis use the network to extract lowdimensional features.In addition,for the difficulty of training and estimation of the total transformation matrix T in the i-vector model,the training method of the improved total transformation matrix T is given.Subsequently,i-vector features characterizing speaker identity information are trained and extracted.Then,for the problem of mismatch between training data and test data in speech recognition system,this paper proposes a speaker recognition method based on i-vector.In order to further improve the system identification performance and reduce the interference of channel noise,an improved channel compensation method is proposed for the i-vector features extracted from the samples,and a Deep Discriminant Training Network(DDTN)model is obtained.At the same time,for the problem that the adaptive effect is not obvious and the recognition performance is poor,an adaptive training method based on DNN acoustic model is given.Finally,this paper uses the GPU accelerated model training in the Kaldi speech recognition platform,and uses the TIMIT and Switchboard corpus to experiment and analyze the adaptive method proposed in this paper.The experimental results show that the proposed method has a lower system recognition error rate than the traditional i-vector feature extraction method and adaptive training method.It proves the rationality and effectiveness of the speaker adaptive method proposed in this paper.