eduzhai > Applied Sciences > Engineering >

Unsupervised Discovery of Recurring Speech Patterns Using Probabilistic Adaptive Metrics

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Unsupervised spoken term discovery (UTD) aims at finding recurring segmentsof speech from a corpus of acoustic speech data. One potential approach to thisproblem is to use dynamic time warping (DTW) to find well-aligning patternsfrom the speech data. However, automatic selection of initial candidatesegments for the DTW-alignment and detection of "sufficiently good " alignmentsamong those require some type of pre-defined criteria, often operationalized asthreshold parameters for pair-wise distance metrics between signalrepresentations. In the existing UTD systems, the optimal hyperparameters maydiffer across datasets, limiting their applicability to new corpora and trulylow-resource scenarios. In this paper, we propose a novel probabilisticapproach to DTW-based UTD named as PDTW. In PDTW, distributionalcharacteristics of the processed corpus are utilized for adaptive evaluation ofalignment quality, thereby enabling systematic discovery of pattern pairs thathave similarity what would be expected by coincidence. We test PDTW on ZeroResource Speech Challenge 2017 datasets as a part of 2020 implementation of thechallenge. The results show that the system performs consistently on all fivetested languages using fixed hyperparameters, clearly outperforming the earlierDTW-based system in terms of coverage of the detected patterns.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×