Authors : Manju Patel and Rajeswari Mukesh
Abstract: Human is the most creative and destructive gift given by god to our nature. In order to show his creativity he invented a revolution able device mobile. As mobile phones have become ubiquitous and basic communications tools. It is most supporting tool for increasing the crime. Specially unauthorized multilingual voice calls or voice messages by nonnative person to break the border security. In order to recognize multilingual native speakers on mobile phones we have designed Multilingual Native Speaker Recognition System (MNSRS). To enhance MNSRS we use Multi Layer Feed Forward Neural Network (MLFFNN) in three stages. In the first stage we remove noise from the corrupted voice signal recorded by auto call recorder in mobile by using Hybrid Adaptive Neuro Fuzzy Inference System (ANFIS). We notice that ANFIS gives a faster convergence rate and improved SNR in poor acoustic environment and also increases MNSRS from 53.7-61.6% with 1950 MFCC values. In the second stage, we fused 1950 MFCC features with 13 LPC and 4 prosody features (pitch, intensity, third and fourth formants). With fused 1967 features recognition percentage of MNSRS increases from 61.6-65.5%. In the third stage, over fitting is the cause for degrading performance of MMNSRS.
Manju Patel and Rajeswari Mukesh, 2016. Multistage Enhancement of Multilingual Native Speaker Recognition in Neural Network on Mobile Data. International Journal of Soft Computing, 11: 140-144.