Speaker identification and localization using shuffled MFCC features and deep learning

Abstract The use of machine learning in automatic speaker identification and localization systems has recently seen significant advances. However, this progress comes at the cost using complex models, computations, increasing number microphone arrays training data. Therefore, work, we propose a new end-to-end model based on simple fully connected deep neural network (FC-DNN) just two input microphones. This can jointly or separately localize identify an active with high accuracy single multi-speaker scenarios by exploiting data augmentation approach. In regard, novel Mel Frequency Cepstral Coefficients (MFCC) feature called Shuffled MFCC (SHMFCC) its variant Difference (DSHMFCC). order to test our approach, analyzed performance proposed features different noise reverberation conditions for scenarios. results show that approach achieves these scenarios, outperforms baseline conventional methods, robustness even small-sized