Publication Details

MGB-3 but system: Low-resource ASR on Egyptian YouTube data

VESELÝ Karel, BASKAR Murali K., DIEZ Sánchez Mireia and BENEŠ Karel. MGB-3 BUT System: Low-resource ASR on Egyptian YOUTUBE data. In: Proceedings of ASRU 2017. Okinawa: IEEE Signal Processing Society, 2017, pp. 368-373. ISBN 978-1-5090-4788-8.
Czech title
MGB-3 BUT Systém: egyptské rozpoznávání řeči s omezenými zdroji
Type
conference paper
Language
english
Authors
Veselý Karel, Ing., Ph.D. (DCGM FIT BUT)
Baskar Murali K. (DCGM FIT BUT)
Diez Sánchez Mireia, M.Sc., Ph.D. (DCGM FIT BUT)
Beneš Karel, Ing. (DCGM FIT BUT)
URL
Keywords

MGB-3, ASR adaptation, low-resource ASR, Egyptian Arabic, diarization

Abstract

This paper presents a series of experiments we performed during our work on the MGB-3 evaluations. We both describe the submitted system, as well as the post-evaluation analysis. Our initial BLSTM-HMM system was trained on 250 hours of MGB-2 data (Al-Jazeera), it was adapted with 5 hours of Egyptian data (YouTube). We included such techniques as diarization, n-gram language model adaptation, speed perturbation of the adaptation data, and the use of all 4 correct references. The 4 references were either used for supervision with a confusion network, or we included each sentence 4x with the transcripts from all the annotators. Then, it was also helpful to blend the augmented MGB-3 adaptation data with 15 hours of MGB-2 data. Although we did not rank with our single system among the best teams in the evaluations, we believe that our analysis will be highly interesting not only for the other MGB-3 challenge participants.

Annotation

This paper presents a series of experiments we performed during our work on the MGB-3 evaluations. We both describe the submitted system, as well as the post-evaluation analysis. Our initial BLSTM-HMM system was trained on 250 hours of MGB-2 data (Al-Jazeera), it was adapted with 5 hours of Egyptian data (YouTube). We included such techniques as diarization, n-gram language model adaptation, speed perturbation of the adaptation data, and the use of all 4 correct references. The 4 references were either used for supervision with a confusion network, or we included each sentence 4x with the transcripts from all the annotators. Then, it was also helpful to blend the augmented MGB-3 adaptation data with 15 hours of MGB-2 data. Although we did not rank with our single system among the best teams in the evaluations, we believe that our analysis will be highly interesting not only for the other MGB-3 challenge participants.

Published
2017
Pages
368-373
Proceedings
Proceedings of ASRU 2017
Conference
2017 IEEE AUTOMATIC SPEECH RECOGNITION AND UNDERSTANDING WORKSHOP (ASRU), Okinawa, JP
ISBN
978-1-5090-4788-8
Publisher
IEEE Signal Processing Society
Place
Okinawa, JP
DOI
UT WoS
000426066100051
EID Scopus
BibTeX
@INPROCEEDINGS{FITPUB11595,
   author = "Karel Vesel\'{y} and K. Murali Baskar and Mireia S\'{a}nchez Diez and Karel Bene\v{s}",
   title = "MGB-3 but system: Low-resource ASR on Egyptian YouTube data",
   pages = "368--373",
   booktitle = "Proceedings of ASRU 2017",
   year = 2017,
   location = "Okinawa, JP",
   publisher = "IEEE Signal Processing Society",
   ISBN = "978-1-5090-4788-8",
   doi = "10.1109/ASRU.2017.8268959",
   language = "english",
   url = "https://www.fit.vut.cz/research/publication/11595"
}
Back to top