| The improvement of the level of science and technology has promoted the people’s yearning for a better life.With the development of Internet,AI and other technologies.Biometric identification technologies such as face,fingerprint,and voiceprint are widely used in daily life.The improvement and intelligence of the home environment are an important manifestation of our pursuit of a better life,and the smart home industry has received extensive attention.The traditional smart home is restricted by inconvenient control methods such as manual adjustment and remote control.Speech recognition control technology has promoted the development of the smart home industry.The management and control of home equipment,access control and other equipment through voice brings convenience to people’s lives.At the same time,the requirements for the security of the home scene and the privacy and confidentiality of the householder are getting higher and higher.Carry out research on smart home system based on voiceprint recognition technology that can assign permissions.The main work is as follows:（1）The speaker recognition technical solution is designed,which is divided into two stages,namely the voice and voiceprint input stage and the recognition stage during actual operation.（2）Aiming at the preprocessing of the speaker’s speech signal and the extraction of voiceprint features in the application scenario of the home system.In terms of voice preprocessing,in order to obtain high-quality features and effectively distinguish user voiceprint information from worthless environmental noise,a threshold comparison method is used,which is an endpoint detection method based on logarithmic spectral distance.At the same time,in order to ensure the quality of voiceprints,an improved spectral subtraction noise reduction process is proposed,which effectively improves the quality of voiceprint information and SNR.In terms of voiceprint feature extraction,voiceprint MFCC parameter features are obtained by means of FFT,spectral line energy acquisition,Mel filtering,logarithmic discrete cosine transform DCT,etc.（3）I In the research work of voiceprint recognition model,two different architecture convolutional networks are proposed: Res Net and Rep VGG as the backbone network of the model,and three text-related voiceprint recognition models for home scenes are designed,namely ResSD,Res-SArc and Rep-SArc models.The Res-SD model uses the traditional cross-entropy loss function to complete the training,and the Res-SArc and Rep-SArc models both use the AAM loss function that maximizes the classification limit in the angular space of feature expression to complete the training.Experiments on the self-built smart home speech database（SMARTHOME＿Speech dataset）show that the accuracy rates of the Res-SArc and Rep-SArc models on the test set have reached 97.19% and 97.90%,respectively.In terms of parameter quantity,the parameter quantity of the Rep-SArc model is 1/3 of the parameter quantity of the ResSArc model.The experimental results verify that the three models proposed in this paper are effective for text-related closed-set recognition tasks.In terms of parameter quantity and accuracy,the Rep-SArc model is more suitable for learning user characteristics of household scenes with class discrimination.（4）In the design of smart home system,according to the needs of users and communication networking,the scheme design is carried out according to the hardware and software technology of the system,and the simple model of the smart home system with voice interaction is completed.Assign permissions result response analysis. |
