Whisper to Neutral Mapping Using I-Vector Space Likelihood and a Cosine Similarity Based Iterative Optimization for Whispered Speaker Verification

Abstract

In this work, we propose an iterative optimization algorithm to learn a feature mapping (FM) from the whispered to neutral speech features. Such an FM can be used to improve the performance of speaker verification (SV) systems when presented with a whispered speech. In one of previous works, the equal error rate (EER) in an SV task has been shown to improve by ~24%. based on an FM network trained using a cosine similarity based loss function over that using a mean squared error based objective function. As the mapped whispered features obtained in this manner may not lie in the trained i-vector space, we, in this work, iteratively optimize the i-vector space likelihood (by updating T-matrix) and a cosine similarity based loss function for learning the parameters of the FM network. The proposed iterative optimization improves the EER by ~26% compared to when the FM network parameters are learned based on only cosine similarity based loss function without any T-matrix update, which is a special case of the proposed iterative optimization.