Improved Extreme Learning Machine Based Accent Identification

Automatic identification of foreign accents has a key role in many speech systems,such as speaker identification,e-learning,telephone banking,voice mailing,voice conversions and immigration screenings,etc.Beside it also play vital role to make existing Automatic Speech Recognition(ASR)system robust.English speakers exhibit dialectal differences or non-native accents on specific features of their speech,and these features can be used to identify the accent or native language of the speaker However,automatic accent identification also faces many difficulties,mainly including that accent characteristics are often mixed with language content,rhythm,environmental noise,and the speaker's own voice characteristics.Complex and nonlinear accent recognition models need to be built.In addition,an accent corpus containing a large number of samples also takes a lot of time and effort.In this thesis,through the study of linguistic pronunciation methods,the pronunciation characteristics of the mother tongue can be effectively obtained,and an improved extreme learning machine algorithm is used to obtain a more authoritative and rich English dialect corpus to realize the binary accent classification and multiple accent classification recognition models,respectively.And got better identification results.This framework proposed the consonant phoneme based Extreme Learning Machine(ELM)identification model for Foreign accent identification based on the different pronunciation of English consonant phonemes by Arab native speakers.Mel-Frequency Cepstrum Coefficients(MFCCs)is taken as acoustic features input and trained with DBN,SVMs and ELM classifiers.ELM classifier showed fast learning,and better performance,based on KFold validation with an accuracy of 88% and standard deviation( = 0.0167),76% by SVM and 64% with DBN classifier respectively.Our proposed ELM model showed 11%,16% accuracy increment respectively over the previous work model by using the same classifier on multiple words based acoustic model to identify regional accents.Mostly,traditional models that are designed for Binary classification reflects good results,also few good results found with Multi-classification,but usually accuracies and model's overall exhibits low performances.In this paper,we proposed Multi-Kernel based Extreme Learning Machine(MKELM)Model,for Foreign Accent Identification using Multi-class classification.Proposed model uses a novel weighted scheme to classify six different native languages including,English,Arabic,Chinese,Korean,French and Spanish.We used MFCCs and prosodic features combination,that are used as input raw feature for models training.Our proposed model for multi-classification gives better accuracy of 82.75% using pairwise weighted Scheme,as a comparison SVM and KELM shows 71%,81.2% respectively.For the traditional multi-classification proposed model yield 66.5% accuracy as a comparison SVM,ANN,LSTM,ELM,MLELM and KELM models shows 39%,21%,27.1%,32%,37%,65% respectively.For the model's performance comparison our proposed model takes less training time 45(Sec)as compared to SVM which takes almost double of our proposed model 92(Sec).Overall performance,accuracy reported for multi-classification reflects the advantages of our proposed model.