Conference paper

VESELÝ Karel, PERALES Carlos Segura, SZŐKE Igor, LUQUE Jordi and ČERNOCKÝ Jan. Lightly supervised vs. semi-supervised training of acoustic model on Luxembourgish for low-resource automatic speech recognition. In: Proceedings of Interspeech 2018. Hyderabad: International Speech Communication Association, 2018, pp. 2883-2887. ISSN 1990-9770. Available from: https://www.isca-speech.org/archive/Interspeech_2018/abstracts/2361.html
Publication language:english
Original title:Lightly supervised vs. semi-supervised training of acoustic model on Luxembourgish for low-resource automatic speech recognition
Title (cs):Trénování akustického modelu lucemburštiny pro automatické rozpoznávání řeči s omezenými zdroji s lehkou supervizí vs. bez supervize
Pages:2883-2887
Proceedings:Proceedings of Interspeech 2018
Conference:Interspeech 2018
Place:Hyderabad, IN
Year:2018
URL:https://www.isca-speech.org/archive/Interspeech_2018/abstracts/2361.html
Journal:Proceedings of Interspeech - on line, Vol. 2018, No. 9, BAIXAS, FR
ISSN:1990-9770
DOI:10.21437/Interspeech.2018-2361
Publisher:International Speech Communication Association
URL:http://www.fit.vutbr.cz/research/groups/speech/publi/2018/vesely_interspeech2018_2361.pdf [PDF]
Keywords
Luxembourgish, call centers, speech recognition, low-resourced ASR, unsupervised training
Annotation
In this work, we focus on exploiting inexpensive data in order to to improve the DNN acoustic model for ASR. We explore two strategies: The first one uses untranscribed data from the target domain. The second one is related to the proper selection of excerpts from imperfectly transcribed out-of-domain public data, as parliamentary speeches. We found out that both approaches lead to similar results, making them equally beneficial for practical use. The Luxembourgish ASR seed system had a 38.8% WER and it improved by roughly 4% absolute, leading to 34.6% for untranscribed and 34.9% for lightlysupervised data. Adding both databases simultaneously led to 34.4% WER, which is only a small improvement. As a secondary research topic, we experiment with semi-supervised state-level minimum Bayes risk (sMBR) training. Nonetheless, for sMBR we saw no improvement from adding the automatically transcribed target data, despite that similar techniques yield good results in the case of cross-entropy (CE) training.
BibTeX:
@INPROCEEDINGS{
   author = {Karel Vesel{\'{y}} and Segura Carlos Perales and Igor
	Sz{\H{o}}ke and Jordi Luque and Jan {\v{C}}ernock{\'{y}}},
   title = {Lightly supervised vs. semi-supervised training of acoustic
	model on Luxembourgish for low-resource automatic speech
	recognition},
   pages = {2883--2887},
   booktitle = {Proceedings of Interspeech 2018},
   journal = {Proceedings of Interspeech - on line},
   volume = {2018},
   number = {9},
   year = {2018},
   location = {Hyderabad, IN},
   publisher = {International Speech Communication Association},
   ISSN = {1990-9770},
   doi = {10.21437/Interspeech.2018-2361},
   language = {english},
   url = {http://www.fit.vutbr.cz/research/view_pub.php?id=11844}
}

Your IPv4 address: 54.224.150.24
Switch to IPv6 connection

DNSSEC [dnssec]