Articulatory Features Based TDNN Model for Spoken Language Recognition

Abstract

In order to improve the performance of the Spoken Language Recognition (SLR) system, we propose an acoustic modeling framework in which the Time Delay Neural Network (TDNN) models long term dependencies between Articulatory Features (AFs). Several experiments were conducted on APSIPA 2017 Oriental Language Recognition(AP17-OLR) database. We compared the AFs based TDNN approach to the Deep Bottleneck (DBN) features based ivector and xvector systems, and the proposed approach provide a 23.10% and 12.87% relative improvement in Equal Error Rate (EER). These results indicate that the proposed approach is beneficial to the SLR task.