Exploring Acoustic Factor Analysis for limited test data speaker verification
Salil Mamodiya, Lav Kumar, Rohan Kumar Das, S. R. Mahadeva Prasanna · 2016
In text independent speaker verification (TI-SV) domain, recently proposed Acoustic Factor Analysis (AFA) model has shown its importance over conventional i-vector based approach. AFA takes into account the redundancies present in the mel frequency cepstral coefficient (MFCC) features. It transforms the features to a lower dimensional space which is much close to speaker subspace. In practical applications duration of the test data is very important for SV task. It may not be always the case that the test data of sufficient duration is provided. Limited data have less phonetic content, that makes TI-SV under limited data a challenging task. Previously using i-vector based approach on MFCC features, it has been proved that performance of the SV task drops as the duration of the test data is reduced. This work attempts to improve performance of SV task for limited duration test utterances using AFA model. A SV system is built based on AFA. A parallel SV system of conventional i-vector based approach using MFCC features is also created. Then SV is carried out for limited test data conditions (≤10 s) on NIST SRE 2003 dataset using both the AFA and the i-vector based approach, AFA showing improved results over the latter case. The systems are then fused at the score level and their combination is found to give significant improvement over baseline performance highlighting importance of AFA based modeling approach for limited test data condition.