Incorporating Local Acoustic Variability Information into Short Duration Speaker Verification

Abstract

State-of-the-art speaker verification systems are based on the total variability model to compactly represent the acoustic space. However, short duration utterances only contain limited phonetic content, potentially resulting in an incomplete representation being captured by the total variability model thus leading to poor speaker verification performance. In this paper, a technique to incorporate component-wise local acoustic variability information into the speaker verification framework is proposed. Specifically, Gaussian Probabilistic Linear Discriminant Analysis (G-PLDA) of the supervector space, with a block diagonal covariance assumption, is used in conjunction with the traditional total variability model. Experimental results obtained using the NIST SRE 2010 dataset show that the incorporation of the proposed method leads to relative improvements of 20.48% and 18.99% in the 3 second condition for male and female speech respectively.

An Improved Deep Embedding Learning Method for Short Duration Speaker Verification

“…And w is the speaker and channel factor with a standard normal distribution N (0, I). However, it is known that for short duration SV, the statistics are not sufficient for reliable i-vector learning, which leads to degraded performance [15]. Furthermore, the generative models obtained via unsupervised learning methods may be improved with discriminative models, e.g.…”