Ml speech depth neural network model

Reference resources

Huawei cloud College

DNN-HMM

Deep neural network - The hidden Markov model （DNN-HMM） utilize DNN Characteristic learning ability and HMM The serialization modeling ability of is used to process speech recognition tasks , In many large-scale missions , Its performance is much better than the traditional GMM-HMM hybrid model .
DNN： Characteristic learning ability , Estimate the probability of observing a feature , A posteriori probability of the predicted state .
HMM： Describe the sequence change of speech signal , Predict the next sequence .

DNN-HMM speech recognition

Acoustic signal use HMM Framework modeling , The generation law of each state uses DNN Replace the original GMM,DNN The output of each cell represents the posterior probability of the state .

CD-DNN-HMM

although GMM-HMM In the past, it has achieved a lot of success , But with the development of deep learning ,DNN Shows that GMM Greater advantages . differ GMM,DNN Context information is introduced （ Front and back feature frame information ）, go by the name of CD-DNN-HMM（Context-Dependent DNN-HMM） Model .

CD-DNN-HMM form

CD-DNN-HMM It's made up of three parts ：DNN（1）,HMM（1）, State prior probability distribution （1）. because CD-DNN-HMM and GMM-HMM Shared factor binding structure , So training CD-DNN-HMM The first step is to use training data to train a GMM-HMM, utilize Viterbi The standard results produced by decoding are used for DNN.

CD-DNN-HMM Performance improvement

（1） Use deeper Neural Networks .（2） Use longer frames as input .（3） Use three factors to model .（4） Improve the annotation quality of training data .（5） Preliminary training （ Shallow DNN）.

DNN Training accelerates

（1） many GPU Back propagation .（2） Asynchronous random gradient descent .（3） Reduce model size .（4） Integrate .

DNN Decoding acceleration

（1） Parallel computing .（2） Sparse networks .（3） Low rank approximation .（4） Multiframe DNN.

DNN because GMM

DNN It's a discriminant model , It is different in itself , It can better distinguish annotation categories .DNN Excellent performance in big data , With the increasing amount of data ,GMM Model in 2000 Performance saturation will occur in about hours , and DNN Can support 10000 hours .DNN It has stronger effect on noise robust, Through noise training ,DNN The recognition performance of the model in complex environment can even exceed that processed by speech enhancement algorithm GMM Model .