Unsupervised Bottleneck Features for Low-Resource Query-by-Example Spoken Term Detection

We propose a framework which ports Dirichlet Gaussian mixture model (DPGMM) based labels to deep neural network (DNN). The DNN is used to extract a low-dimensional unsupervised speech representation, named as unsupervised bottleneck feature (uBNF), which captures considerable information for sound cluster discrimination. We investigate the performance of uBNF in query-by-example spoken term detection (QbE-STD) on the TIMIT English speech corpus. Our uBNF performs comparably with the cross-lingual bottleneck features (M-BNF) extracted from a DNN trained using 171 hours of transcribed telephone speech in another language (Mandarin Chinese). With the score fusion of uBNF and M-BNF, we gain about 10% relative improvement in term of mean average precision (MAP) comparing with M-BNF. We also studied the performance of the framework with different input features and different lengths of temporal context.