Pairwise learning using multi-lingual bottleneck features for low-resource query-by-example spoken term detection.