Multilingual Bottle-Neck Feature Learning From Untranscribed Speech For Track 1 In Zerospeech2017 (System 2 -- With Vtln)

We investigate the extraction of bottle-neck features (BNFs) for multiple languages without access to manual transcription. Multilingual BNFs are derived from a multi-task learning deep neural network which is trained with unsupervised phoneme-like labels. The unsupervised phoneme-like labels are obtained from language-dependent Dirichlet process Gaussian mixture models separately trained on untranscribed speech of multiple languages. In this version, the input MFCC for DPGMM is processed with VTLN.