Research On Key Technologies Of Voiceprint Recognition In Household Scenarios

As a core part of China's Smart-City strategy,smart home is a safe and reliable access control system that is through real-time and accurate personal identification,to achieve family member access and foreign personal unaccess.Voiceprint recognition is the best choice for biometric identification of home access control due to its outstanding advantages in safety,convenience,interactive naturalness,and few economies.The voiceprint recognition of home access control scenes has the characteristics of small samples and short speeches.How to achieve accurate voiceprint recognition and ensure the absolute security of the home environment is a scientific problem to be solved urgently.Aiming at these two scientific problems,based on the ideas of parallel and series systems in engineering,this thesis proposes a multi-feature fusion voiceprint recognition algorithm and a CNNfeature matching algorithm.The specific research contents and main results are as follow:?.The theories and implementations of each step of the voiceprint recognition system,for instance,LPC,LPCC,MFCC,PLP and template matching,GMM,SVM,are studied in detail,which have laid the foundation of proposing the two new algorithms;?.In response to the scientific problem of small sample phrase sounds,a multi-feature fusion voiceprint recognition algorithm is innovatively proposed.The three sub-models are trained separately and then fused in parallel to realize the full use of limited information.A voiceprint feature correlation evaluation method is constructed.A two-stage training method is proposed,and is verified by experiment.The results show that the multi-feature algorithm has a better classification capability on voiceprint than LPC,LPCC,and MFCC sub-models,and achieves accurate classification of short voice with a accuracy rate of 98%;?.In view of the high security requirements of home access control scenarios,a joint CNN and feature matching algorithm is proposed.By connecting the CNN model and the feature matching model,the FAR of the voiceprint recognition system is reduced.CNN classification based on the spectrogram is modeled and trained,and a feature matching model for second judgement is established,and experimental verification is carried out.The results show that the FAR of the joint algorithm is reduced by 75% and 94.76% compared to the pure CNN and feature matching model,respectively,namely household safety is secured.