An Extension of the Slovak Broadcast News Corpus based on Semi-Automatic Annotation

An Extension of the Slovak Broadcast News Corpus based on Semi-Automatic Annotation Peter Viszlay, J´an Staˇs, Tom´aˇs Koct´ur, Martin Lojka, Jozef Juh´ar Department of Electronics and Multimedia Communications Faculty of Electrical Engineering and Informatics, Technical University of Koˇsice Park Komensk´eho 13, 042 00 Koˇsice, Slovak Republic {peter.viszlay, jan.stas, tomas.koctur, martin.lojka, jozef.juhar}@tuke.sk Abstract In this paper, we introduce an extension of our previously released TUKE-BNews-SK corpus based on a semi-automatic annotation scheme. It firstly relies on the automatic transcription of the BN data performed by our Slovak large vocabulary continuous speech recognition system. The generated hypotheses are then manually corrected and completed by trained human annotators. The corpus is composed of 25 hours of fully-annotated spontaneous and prepared speech. In addition, we have acquired 900 hours of another BN data, part of which we plan to annotate semi-automatically. We present a preliminary corpus evaluation that gives very promising results. Keywords: automatic subtitling, broadcast news, semi-automatic annotation, speech-to-text transcription 1. Motivation and Previous Work The initial research related to Broadcast News (BN) tran- scription for the Slovak language started in recent years, when the BN data collection was performed. This activity resulted in the first Slovak TUKE-BNews-SK corpus, intro- duced in (Pleva and Juh´ar, 2014). It consists of 265 hours of spontaneous speech, dialogues and live coverage with dif- ferent background conditions. It was manually annotated, because there were no other available resources supporting the annotation process at that time. It is still the only Slo- vak annotated corpus for building knowledge sources for the BN transcription task. In this paper, we introduce a new extension of the TUKE- BNews-SK corpus, which stems from the current needs of our laboratory, and whose creation was partially initiated by demand in Slovak research. The reason for building the corpus lies in the fact that the modern and leading trends in building resources for large vocabulary continuous speech recognition (LVCSR) applications focus on fully-automatic annotation of speech (Wessel and Ney, 2005; Novotney et al., 2009), without any additional human annotation effort. This fact motivated us to follow this trend. Unfortunately, we still do not have sufficient acoustic re- sources to perform automatic annotation, because the pre- vious corpus did not provide enough data to build accurate acoustic models (AMs) for that purpose. With our increas- ing effort to build more precise and robust speech transcrip- tion system for error free speech transcription, the need for more data of high-quality become stronger and resulted re- cently in a cooperation with one of the Slovak commer- cial TV broadcasters. Moreover, with increasing demand driven by the growing population of deaf and hearing im- paired people (approx. 5 − 7% of Slovak population), we started recently with this broadcaster to support more TV broadcasting with Slovak closed captions generated auto- matically through modern speech recognition technologies. The above mentioned facts really motivated us to create a new corpus, which would meet the conditions and re- quirements for building modern LVCSR systems suitable for automatic annotation purposes. In the …

Content truncated to 500 words