Study On Garbage Classification Collection System Based On Speech Recognition Technology

At present,the traditional manual classification method is still used for garbage classification.Which not only makes the classification efficiency low,but also prone to classification errors.So as to settle the difficulty of low classification accuracy resulting from manual garbage classification,speech recognition technology is applied to garbage classification.The garbage classification and collection system based on speech recognition technology is studied through algorithms,models,and embedded software and hardware.Related algorithms in speech recognition technology are focused on study.First,the improvement strategies are proposed after the traditional algorithms have been researched.Then,the improved algorithms are compared with the traditional algorithms.Finally,according to the required conditions,a software system and a hardware system are designed.The main work of the thesis is summarized as follows:(1)Function analysis and overall scheme construction of system.Starting from the system function requirements,the overall system functions are analyzed on the basis of the four basic principles of advancement,ease of use,system,and economy.According to the system requirements,speech data acquisition,speech signal preprocessing,speech signal training,speech recognition,automatic opening and closing of the trash can lid,and overflow detection of the trash can need to be included.Finally,the overall scheme of the system is built according to the requirements and analysis.(2)Study on Related Algorithms of Speech Signal Preprocessing.First,the input speech signal is preprocessed.Then,speech enhancement algorithm,endpoint detection algorithm and feature extraction algorithm are studied,the traditional algorithms are improved after analysis,and the traditional algorithms are used for comparative with the improved algorithms.After evaluation,compared with the other algorithms,the improved speech enhancement algorithm Perceptual Evaluation of Speech Quality(PESQ)increased by 14.71%～45.70%,the Logarithmic Spectral Distance(LSD)has decreased by 18.14%～25.47%,and the Source Distortion Rate(SDR)by From-5.00～11.00 to 2.00～14.00;When the signal-to-noise ratio is lower than-10dB,the detection accuracy of the speech endpoint detection algorithm is more than 85%,and the average detection time is shortened to 1/3 of the traditional algorithms;For speech signal feature extraction algorithms,when the signal-to-noise ratio of the algorithm is-10dB,compared with the traditional algorithms,the speech recognition rate is increased by an average of 24.05%.In terms of time performance,the average training time is reduced by 23.20%and the average recognition time is reduced by 32.37%.(3)Speech acoustic model construction.First,traditional speech signal models are studied,the current mainstream speech signal training models are focused on study,then a deep neural network model is established.On this basis,the adaptive deep neural network model is studied.Speech recognition in complex environment is realized by improving regularization adaptive criterion and output layer activation function.Test experiments by superimposing background noise on multiple speech data sets,the results show that,compared with the currently popular GMM-HMM and traditional DNN speech acoustic models,the recognition word error rate is reduced by 5.15%,3.11%.Then,the acoustic model of convolutional neural network is studied,and a three-layer optimized convolutional neural network speech recognition is designed.Through a variety of evaluation indicators to compare the model before and after the improvement,the results display,compared with the comparison algorithms,the average recognition error rate of the Chinese speech data set is reduced by 22.05%,and the average recognition error rate of the English speech data set is reduced by 20.27%,which is 40%less than the loss of the traditional convolutional neural network model.(4)Hardware selection and design of system.Through the selection and analysis of the microprocessor,the ARM9 series chip STM32 is used as the main control module for the processor.According to the actual scenarios and requirements of the system,the main controller module,speech recognition module,motor drive module,ultrasonic distance measurement module,and display circuit module are designed in detail.(5)Software design and overall function test of system.First,the software development platform is established,Keil5 is used as a program development tool.Next,the Linux kernel is cut and transplanted.Then,the main control center,speech recognition module and motor drive module are designed.The speech database is established according to the system requirements.Finally,the system functions are tested as a whole.The garbage classification collection system based on speech recognition is studied,it can be achieved without remembering the type of garbage,the following functions can be realized by reporting the name of the garbage:automatic voice recognition,automatic opening and closing of the lid of the garbage bin,and detection of garbage capacity.The system can reduce strength of human memory and also reduce the error rate of manual classification.Speech recognition related algorithms are focused on study in this thesis,after the traditional algorithms are analyzed,the optimized algorithms are designed.Evaluation by indicators,the optimized algorithms performance have been improved.