Research On Speech Recognition System For Fuzzy Instruction

With the deepening of industrial automation,voice manipulation of industrial robots has become a mainstream area of research today.However,in the application of voice interaction and voice operation of industrial robots,huge noises are generated in the working environment of robots,and these noises can greatly affect the recognition effect.In addition,some workers or engineers do not strictly comply with the requirements of the workbook when issuing voice commands and use some colloquial words in the voice commands.In all,those factors can lead to poor recognition of commands in industrial robots.In this thesis,an algorithm model is proposed to solve this problem.The model consists of two parts,namely the speech enhancement and the speech keyword recognition.The former is used for the decrease of influences brought by the environmental noise on the command recognition.The involved Res-Unet network,compared with the traditional U-Net network framework,incorporates a residual network in the encoder and the decoder,which increases the convergence and complexity of the network and improves its expressiveness.Through experiments in noisy environments of industrial robots with PESQ and STOI as evaluation criteria,it is found that the performance of Res-Unet network is better than other speech enhancement methods,which bears great significance in speech enhancement in industrial noisy environment.The latter,the speech keyword recognition,is employed to extract keywords in speech commands,with which it can complete the manipulation of robot before fully understanding the whole command.This method can effectively reduce negative effects of speech commands with ambiguous semantics on recognition.It uses CTC as the loss function and applys CNN-Bi GRU network,which is selected by comparing recognition accuracy of eight CNN-RNN networks for the same dataset.In the Libri Speech speech dataset,CNN-Bi GRU network achieves the best result,with 97% recognition accuracy for specific keywords.Based on those two algorithmic models,the recognition problem of fuzzy commands can be well solved,which can make industrial robots better manipulated by voice commands.Therefore,the algorithm model in this paper is of great importance for the development of industrial automation.Finally,this paper designs and implements a fuzzy command recognition system,using a microphone for real-time audio acquisition and two algorithmic models to process the audio and obtain command recognition results.After experiments,it is shown that the system has a good recognition accuracy.