Voice Recognition Based On Recurrent Neural Network

Voice is one of the most common information carriers and can support the transfer of information.In addition,compared with other types of sensors,acoustic sensors do better in concealment,and the cost is very low.In the case of electromagnetic interference,the ability to resist external interference is stronger than other types of sensors.Today,the recognition of sound signals is widely used in many fields.Therefore,it is of great significance to study the detection of sound targets.The two modules of feature extraction and classification are the core parts of research on sound detection.In the traditional mode,features are generally extracted manually,which requires high personal experience.In addition,complex sound features in certain environments are difficult to extract manually.This will make it impossible to construct a classifier and classify complex sounds.Deep learning algorithm,as an intelligent perception algorithm,can efficiently mine category attributes and deep features.Based on the theoretical foundation of deep learning,this paper proposes the use of different neural network algorithms in deep learning to construct a classifier for the detection of sound targets.Based on the two main modules of sound feature extraction,classification and recognition,this paper designed two different deep learning algorithms for convolutional neural network(CNN)and recurrent neural network(RNN)suitable for processing audio signals.Use these two algorithms to realize the learning of sound features and achieve correct classification.The innovation of this paper mainly has the following four points:1.Using deep learning model to design engine noise recognition system.2.Use convolutional neural network to extract the sound features of multiple noise sources and multiple durations.3.Optimized in training network and training method.4.Designed different kinds of deep learning network models for experimental comparison to get the best one.Based on the above content,the experimental results are expected to be applied to adaptive multi-target sound detection.The content of the paper includes the following aspects:1.Introduce the research background of the subject,briefly describe the application field of sound detection and the research status in this area.2.Introduce the relevant theoretical basis of sound detection,as wellas some basic knowledge of deep learning related algorithms,CNN and RNN including long and short time memory model(LSTM)and gated cycle unit(GRU).3.An engine noise recognition system is designed based on the deep learning network.The overall architecture of the system and the design process are introduced.It focuses on the multi-time MFCC feature extraction method,as well as the structure and training methods of different deep learning training networks.4.Compare the results of training using different deep learning networks and give an analysis and get conclusions.5.Summarize the work of the paper and discuss the unresolved problems and the direction of future development.