Research And Implementation Of Telephone Classification Method Based On Audio Features

In recent years,the telephone classification technology based on signaling and voice content has achieved extensive research results,but it is difficult to effectively apply in the telephone classification problems with incomplete signaling or weak relation of voice content.The problem of telephone classification is very important in many application scenarios.This thesis studies this problem and proposes a method of telephone classification based on audio features.This thesis proposes a VoIP telephone classification method based on audio statistical features.This method extracts and calculates audio features based on statistical method,which can distinguish VoIP and telecom phones.Then it uses the random forest and LightGBM based on decision tree to realize feature filtering,and uses the features with high importance coefficient for multi model training to realize the classification of VoIP and telecom phones.This thesis proposes a call forwarding telephone classification method based on spectrogram training of auto-encoder.This method extracts audio features that can be used to distinguish the call forwarding phones from the ordinary ones by hidden features obtained from de-noise auto-encoder which reflect noise characteristics in addition to features calculated from statistical method.This method implements call forwarding telephone classification by feature combination,feature selection and multi model training.This thesis proposes a telephone source classification method based on sequence feature trained from neural network.This method uses MFCC,PLP features and their difference features to train convolutional neural network for language recognition,and then trains use multiple audio sequence features to train convolutional neural network and GRU network to obtain hidden features,and realizes geographic location classification by combining statistical features and hidden features trained by deep auto-encoder from spectrogram.The author collected 8000 VoIP phones,6000 call forwarding phones,7000 ordinary phones and 10000 overseas phones.In VoIP telephone classification,F1 score reaches the highest 90.15%;In call forwarding telephone classification,F1 score reaches the highest 91.88%;In phone source classification,with MFCC and PLP features used in language recognition part,Japanese-Chinese and English-Chinese classification achieve ideal results.In the part of geographical location classification,with features trained from convolutional neural network,for country A,the highest F1 score is 91.06%,country B,79.71%,and country C,the highest F1 score is 74.26%.Experimental results show that the telephone classification based on audio features is effective.This thesis also implements a telephone classification system based on audio features.The user can classify the telephone by importing test audio and specifying the classification module.The system test results show that the system can respond to the user's request and return the correct results.