Research Of Speech Recognition Based On Kaldi

With the rapid development of artificial intelligence technology,the drawbacks of traditional human-computer interaction methods are gradually exposed,and people's demand for new interactive products becomes more urgent.Simultaneously,the development of big data,ripple effects,and deep neural networks has brought a burst of voice technology,and voice technology has entered a new era.The application of voice technology in interactive means and the development of voice interaction systems have become new research hotspots.Kaldi is a comprehensive object-oriented toolkit written in C++ and was developed by Dr.Dan Povey from the former Microsoft Research Institute and the University of the BUT in the Czech Republic.It is powerful and most of the mainstream models,algorithms and data at the present stage are supported.New technologies,new ideas and new ideas can be acquired more easier for its open source features.Its multi-platform compatibility features are suitable for development work.Modeling,training and decoding of hybrid Gaussian Hidden Horse Model and neural network Hidden Horse Model are studied in this thesis.The algorithm,data processing methods and process acceleration techniques used in the model are analyzed in depth;the Kaldi source code is deeply studied and analyzed.Based on its internal functions and implementation mechanism,performance testing and comparison of hybrid Gaussian Hidden Horse model and neural network based on Kaldi was conducted,and the superiority of the neural network Hidden Horse model to the hybrid Gaussian Hidden Horse model was recognized.Aiming at the problem of too long training time caused by training over fitting in neural network training,a new function as an activation function are proposed in this paper proposes.Experiments show that the function has achieved effective results in solving gradient dispersion problems,and the training time has been reduced to some extent.It is believed that if the function is further studied and optimized,the hardware requirements of neural network training can be reduced and the further development of neural network technology can be promoted.