Design And Implementation Of Real-time Civil Aviation Speech Recognition Algorithm Based On Deep Learning

The aim of speech recognition technology is to transform speech into text.It can be used to translation,education,military,medical and other fields to boost work efficiency.This thesis mainly researches the real-time civil aviation speech recognition method based on deep learning,which transform between speech features and corresponding text,and implements the real-time recognition of civil aviation commands.After consider through the devise and modeling of existing algorithms in-depth analysis and discover that there are several issues of these algorithms: First,CNN-based models are more suitable for real-time speech recognition because they only rely on limited context for prediction.Since the model greater depth,and for the sake of receive a greater range of receptive fields at a limited depth,using a greater convolution kernel,this leads to a large amount of model parameters and high calculated amount,and reduces the processing speed of the model.It is hard to satisfy real-time demand;the second is that most of the current ASR models are implemented based on DNN,and the performance of this type of model depends on a large-scale annotation corpus.The end-to-end model is even more so,and in the civil aviation field The cost of data acquisition is very high,so it is difficult to obtain large-scale civil aviation annotation data,which cause it hard for end-to-end models to realize good property.Third,at large of the available corpora are collected in a relatively quiet environment.Use this The data-trained model tends to be very poor when transcribing noise-containing speech,and even hard to use.This thesis researches the above questions and presents relevant terms of settlement.The primary dedication are as follows:(1)Propose a real-time civil aviation speech recognition algorithm based on transfer learning.Aiming at problem 1,separable convolution is applied to decrease the model arguments and the quantity of numeration.In response to question 2,the thought of transfer learning is imported.Firstly,use a large-scale open corpus to train the model’s transcription ability.After the model converges,use a small-scale civil aviation corpus for domain adaptation,so that the model can achieve better results in the civil aviation field.Effective,and has generalization in the open domain.laboratory finding indicate that the suggested method has greater recognition result and faster conducting speed in the field of civil aviation.(2)Propose a deep loop network speech enhancement algorithm based on Attention.For question three,considering the capacity of speech enhancement to boost speech sharpness,it can be used in the front end of speech recognition.Deep recurrent networks are used to enhance speech,and the attention mechanism is applied to acquire contextual knowledge to further increase the enhancement result.Through contrast with the basic method,the availability of this method is confirm.(3)Based on the previously proposed end-to-end method based on separable convolution and transfer learning,a real-time speech recognition system for civil aviation is designed and implemented.The voice recognition requirements in the civil aviation field were analyzed,and then the voice recognition system was devised and realized,and the system was assessed in the end.