Best fusing of acoustic and prosodic features: Application to speaker recognition

摘要

This paper assesses two popular speaker features prosodic and cepstral coefficient. We compare the performance of individual features and features combined via PCA and LDA. The identification process can be performed both in the temporal and cepstral domains. The result show that the reduced sets of the composite LDA feature allow more robust estimates for the model parameters and improve the Recognition Rate (RR), as well as reducing the size which is crucial for real-time speaker recognition application using low-resource devices.