Unsupervised Discovery of Recurring Speech Patterns Using Probabilistic Adaptive Metrics

Abstract

Unsupervised spoken term discovery (UTD) aims at finding recurring segments of speech from a corpus of acoustic speech data. One potential approach to this problem is to use dynamic time warping (DTW) to find well-aligning patterns from the speech data. However, automatic selection of initial candidate segments for the DTW-alignment and detection of "sufficiently good" alignments among those require some type of predefined criteria, often operationalized as threshold parameters for pair-wise distance metrics between signal representations. In the existing UTD systems, the optimal hyperparameters may differ across datasets, limiting their applicability to new corpora and truly low-resource scenarios. In this paper, we propose a novel probabilistic approach to DTW-based UTD named as PDTW. In PDTW, distributional characteristics of the processed corpus are utilized for adaptive evaluation of alignment quality, thereby enabling systematic discovery of pattern pairs that have similarity what would be expected by coincidence. We test PDTW on Zero Resource Speech Challenge 2017 datasets as a part of 2020 implementation of the challenge. The results show that the system performs consistently on all five tested languages using fixed hyperparameters, clearly outperforming the earlier DTW-based system in terms of coverage of the detected patterns.

Word Segmentation on Discovered Phone Units with Dynamic Programming and Self-Supervised Scoring

“…The qualitative analyses in this paper revealed a bias towards segmenting shorter (often filler) words. One approach would be to incorporate top-down information from a sparse term discovery system tailored towards discovering recurring but longer words [52], [53], [55], which could be used to bootstrap segmentation. Another approach would be to incorporate information from another modality; we know that infants have access to cross-situational cues from different modalities that can aid word learning [56].…”

Do Infants Really Learn Phonetic Categories?

“…There have also been recent proposals for more domain-specific perceptual space learning methods that rely on a noisy top-down signal provided by knowledge of some word-like units (Kamper et al, 2015 ; Renshaw et al, 2015 ; Riad et al, 2018 ; Thiollière et al, 2015 ). These units can be found by searching for stretches of speech that form similar pairs or clusters, without any knowledge of phones (Jansen & Van Durme, 2011 ; McInnes & Goldwater, 2011 ; Park & Glass, 2008 ; Räsänen & Blandon, 2020 ). Assuming that the clusters represent different instances of the same word, the learner can then adjust its current representation of the low-level speech features to make these instances even closer together in perceptual space.…”

The Zero Resource Speech Challenge 2020: Discovering discrete subword and word units

“…Two teams, indicated in Figure 2 as B [23] (JHU) and R [24] (Tampere), submitted two systems each. The edge of the grey region in Figure 2e shows the empirical tradeoff previously observed between having high quality matching (low NED) and exhaustively analysing the corpus (high coverage).…”