Publication Details
BUT QUESST 2014 System Description
Skácel Miroslav, Ing. (DCGM FIT BUT)
Burget Lukáš, doc. Ing., Ph.D. (DCGM FIT BUT)
query-by-example, low resourced, neural network, posterior features, bottle-neck features, DTW, AKWS
The paper describes system boild for QUESST 2014 evaluation - query-by-example search in speech. The system is based on DTW and AKWS techniques. It uses posterior and bottle-neck features.
The primary system we submitted was composed of 11 subsystems as the required run. 3 subsystems are based on Acoustic Keyword Spotting (AKWS) and 8 on Dynamic Time Warping (DTW). The AKWS systems were based only on phoneme posteriors while the DTW subsystems were based on both phoneme posteriors and Bottle-Neck features (BN) as input. The underlying phoneme posterior estimators / bottle-neck feature extractors were both in-language (Czech) and out-of-language (other 4 languages). We also performed experiments on T1/T2/T3 types of query, system calibration and fusion based on binary logistic regression.
@INPROCEEDINGS{FITPUB10800, author = "Igor Sz\H{o}ke and Miroslav Sk\'{a}cel and Luk\'{a}\v{s} Burget", title = "BUT QUESST 2014 System Description", pages = "1--2", booktitle = "CEUR Workshop Proceedings", journal = "CEUR Workshop Proceedings", volume = 2014, number = 1263, year = 2014, location = "Barcelona, ES", publisher = "CEUR-WS.org", ISSN = "1613-0073", language = "english", url = "https://www.fit.vut.cz/research/publication/10800" }