Joint Learning of Distance Metric and Query Model for Posteriorgram-Based Keyword Search

An Empirical Evaluation of DTW Subsampling Methods for Keyword Search

“…Figure 3 shows the impact of duration statistics on query construction for the baseline system (note that this figure, unlike Figure 2, includes all terms, and not just OOV terms). The method described in [14] requires that the query sequence be generated by repeating each phoneme representation before concatenating them. This ensures that the constructed queries are approximately as long as they are in the document.…”

“…posteriorgrams from a DNN), learning a mapping from phonemes to vectors in that space, and then aligning such representations so that, in effect, the search becomes QBE. While this method outperforms proxy keywords [14] and, as we'll show, subword OOV retrieval, its practicality is impeded by the computational cost of the DTW algorithm, which is linear (memory and CPU) in the lengths of both the query and document sequences.…”

“…Lexicon expansion methods [9,10] have also been applied, although the effectiveness of such methods naturally depends on how well they are able to preemptively cover the OOV terms without knowing them. DTW has also been used been used for vocabulary independent term retrieval in the context of both query-byexample (QBE) [11,12,13] and keyword search [14,15].…”

“…Moreover, after compression, we maintain a high recall rate as evidenced by the average supremum term weighted value (STWV) of 0.7774, leaving room for search at finer resolution if needed. In text-based KWS, following the method of [14], the vectors of the query sequence come from a finite alphabet, Q = {q 1 , q 2 , . .…”

Collaborative similarity metric learning for face recognition in the wild

“…In this paper, we propose a Siamese neural network‐based SML methodology to facilitate joint learning of a similarity metric that exploits multiple feature representations of the input images. The key contributions of the paper are as follows: • The sigma distance that was recently proposed for speech features in a keyword search task [29], is utilised for face images and compared with the other DML/SML techniques in the literature. • Two COSiM methodologies are proposed and discussed both mathematically and empirically. For experiments, two image representations: (i) scale‐invariant feature transform (SIFT) and (ii) local binary pattern (LBP) are combined and the SML networks are trained jointly to investigate as a means of early fusion.…”