Research On The Key Technology Of Command Word Recognition And Its Application In Respiratory Sound Classification System

The command word recognition,as a method of voice interaction,has become one of key technologies with the development of artificial intelligence technology and the popularization of intelligent equipment.It is also an urgent problem to be solved that keep the robustness of command word recognition under the noise interference situations,which also gotten the academic and industrial attentions.In this paper,an effective solution was proposed to realize the command word recognition in noisy environment.And the key technologies of command word recognition were also utilized on the respiratory sounds recognition domain.This paper mainly focuses on the following three parts:1.In order to ensure the accuracy of command word recognition in noisy environment,a mixed conditions training program is proposed in this paper.Two command word recognition systems are built based on HMM-ML and DNN.HMM.And the experiments of noisy command word recognition are completed using the database of King-ASR-L-057 and our self-made command word database.The experiments show that higher recognition accuracy can be obtained using the mixed training scheme of DNN-HMM under various noisy situations.2.A MFCC fixed-point implementation scheme is proposed to decrease the requirements of system hardware resource.The recognition results of fixed-point and float-point MFCC are compared using the clean command word database based on HMM-ML framework.Then the accuracy of 13 and 39 dimensional fixed MFCCs can achieve 92.12%and 98.40%respectively.At the same time,the command word recognition work can be completed under a certain degree of noisy circumstances,which also verifies the correctness of the fixed-point MFCC extraction process and results.3.In order to solve the respiratory sounds recognition problems,the respiratory sound recognition system are built based on HMM-ML?HMM-MCE and DNN-HMM algorithm respectively,referring the key technologies of command word recognition.A certain scale of respiratory sound database is established by using self-made electronic stethoscope.After finished the normal or abnormal two kinds recognition experiments,the best recognition accuracy of 95.59%can be obtained using the DNN-HMM respiratory sound recognition system.