Study of the Effect of I-vector Modeling on Short and Mismatch Utterance Duration for Speaker Verification

摘要

It is well known that state-of-the-art speaker verification systems using the i-vector concept perform well when target speakers training and test utterances have the same condition: long-long as per NIST evaluation. In practice, real-life applications impose strong constraints on the amount of data that can be used in training target and test speaker models. Since speaker verification systems based on the i-vector approach need to estimate some statistical parameters, the aim of this paper is to explore methods to train statistical parameters of the classical i-vector system when target speakers are trained and tested on mismatched data durations. Experimental results are shown on NIST 2008 SRE for various durations of target training and test speech segments ranging from long to very short, such as full (average 2.5 minutes), 5 seconds and 10 seconds.