Research And Implementation Of Speech Recognition Algorithm Based On HMM And DNN

In the past 2016,artificial intelligence,virtual reality,wearable equipment has become the forefront of technology industry.These studies are inevitably needs to interact with the computer.Compare with the keyboard and mouse,voice is more efficient,and voice can expressing complex emotional,the interactive experience has greatly improved.Therefore,speech recognition technology is bound to be most convenient way of human-computer interaction to widely use.For a long time,the modeling of acoustic models in speech recognition is based on the GMM-HMM model.This model has reliable accuracy and mature EM algorithm for model parameter training.Therefore,GMM-HMM model is widely used in speech recognition.However,the GMM model belongs to the shallow model,the modeling ability is obviously insufficient with the increase of the data volume.Deep neural network(DNN)has become a hotspot in the field of speech recognition because of its better modeling and learning ability for complex data.In this paper,the recognition algorithm based on HMM model and DNN model is studied,and the advantages and disadvantages of the two models are analyzed.This paper has done the following work:(1)This paper studies the speech recognition algorithm based on HMM,and constructs a robot control command speech recognition system using CMUSphinx speech recognition platform.The speech signal of ten control commands of the robot is trained to obtain the language model and the acoustic model.The experimental results show that the average word error rate of the system is 7.1%,and it has a good recognition effect and has a high recognition rate in small vocabulary recognition system.(2)In view of the shortcomings of HMM model,this paper studies the deep belief network(DBN)in deep neural network,and uses Kaldi speech recognition tool to achieve the construction of large vocabulary Chinese continuous speech recognition system.DNN acoustic model trained by Chinese open source speech database named THCHS30.The experimental results show that the word error rate of DNN model is 5.79% less than the triphone model,and the DNN model has better recognition effect in the large vocabulary recognition system.At the same time,this paper uses Kaldi to train the TIMIT speech database to get a large vocabulary English speech recognition system,this paper achieves a better recognition rate.(3)Noise interference has always been a difficulty in speech recognition.In the process of training Chinese acoustic model using Kaldi,DNN training is carried out by adding white noise,car background noise,and coffee background noise in training and testing voice.And compared with a variety of models proved that the DAE model has better effect in learning low-dimensional representation and can be used to restore the input of noise damage.