Investigating the Stacked Phonetic Bottleneck Feature for Speaker Verification with Short Voice Commands

Text-dependent speaker verification (SV) with short voice command (SV-SVC) has increasing demand in many applications. Different from conventional SV, SV-SVC usually uses short fixed voice commands for user-friendly purpose, which causes technical challenges compared with conventional text-dependent SV using fixed phrases (SV-FP). Research results show that the mainstream SV techniques are not able to provide good performance for SV-SVC tasks since they suffer from strongly lexical-overlapping and short utterance length problems. In this paper, we propose to fully explore the acoustic features and contextual information of the phonetic units to obtain better speaker-utterance related information representation for i-vector based SV-SVC systems. Specifically, instead of using MFCC only, the frame-based phonetic bottleneck (PBN) feature extracted from a phonetic bottleneck neural network (PBNN), the stacked phonetic bottleneck (SBN) feature, the cascaded feature of PBN and MFCC, the cascaded feature of SBN and MFCC (SBNF+MFCC) are extracted for developing i-vector based SV-SVC systems. Intensive experiments on the benchmark database RSR2015 have been conducted to evaluate the performance of our proposed ivector SV-SVC systems. It is encouraged that the contextual information learnt from stacked PBNN does help and proposed ivector SV-SVC system with (SBNF+MFCC) outperforms under experimental conditions.