Study On Abnormal Sound Recognition Method Based On UBM And PLDA

Abnormal sound usually refers to a non-voice sound suddenly appearing in a quiet environment,for example,broken glass sounds,infant crying,explosion sounds,knocking sounds,etc.It is widely used in intelligent monitoring,scene recognition,safety monitoring and other occasions.At present,research on abnormal sound monitoring and recognition technique has made certain progress and has been applied in practice.However,the recognition rate is relatively low and robust performance is poor,and it cannot meet actual needs.This thesis studies the abnormal sound monitoring and recognition problem,and improves the recognition performance by applying the uniform background model and PLDA model.The main tasks are as follows:(1)This thesis briefly introduces the abnormal sound feature extraction methods,and the traditional classification models for abnormal sound recognition.(2)The Universal Background Model(UBM)recognition system is introduced in detail.The thesis studies the universal background model training parameters algorithm and the maximum adaptive criterion.Experiments are conducted to verify the effect of the Gaussian model's classification components and the dimensionality of the Mel-Frequency Cepstrum Coefficient(MFCC)in the universal background model on the experimental performance.Besides,the advantages and disadvantages of the two models are also compared.(3)Construct the abnormal sound recognition system based on PLDA model.The extraction algorithm of identity vectors and PLDA model are described and the total variation space matrix is estimated.Moreover,the dimensionality of the identity vector is reduced using the space matrix and factor analysis techniques,and the PLDA model is used to calculate the scoring system.In simulation experiments,the effects of different dimensions of the identity vector and the PLDA factor dimension on the recognition rate are compared.