Large Corpus of Czech Parliament Plenary Hearings

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6363–6367 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6363 Large Corpus of Czech Parliament Plenary Hearings Jon´aˇs Kratochv´ıl, Peter Pol´ak, Ondˇrej Bojar Charles University Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics surname@ufal.mff.cuni.cz Abstract We present a large corpus of Czech parliament plenary sessions. The corpus consists of approximately 444 hours of speech data and corresponding text transcriptions. The whole corpus is segmented to short audio snippets making it suitable for both training and evaluation of automatic speech recognition (ASR) systems. The source language of the corpus is Czech, which makes it a valuable resource for future research as only a few public datasets are available for the Czech language. We complement the data release with experiments of two baseline ASR systems trained on the presented data: the more traditional approach implemented in the Kaldi ASR toolkit which combines hidden Markov models and deep neural networks and a modern ASR architecture implemented in Jasper toolkit which uses deep neural networks in an end-to-end fashion. Keywords: Czech, automatic speech recognition, dataset, speech corpus 1. Introduction The field of automatic speech recognition (ASR) has been recently undergoing a methodological shift from hybrid model architectures towards end-to-end systems. These mostly neural-network-based architectures have recently been gaining momentum in their popularity and obtained state-of-the-art results on multiple popular speech corpora, e.g. Han et al. (2019) on test-clean set from LibriSpeech (Panayotov et al., 2015). One particular downside of the end-to-end approaches is their requirement of an extensive amount of training data in order to produce competitive results in comparison with more traditional hybrid architectures. At the same time, there has been active research in the field of fluent speech reconstruction (Fitzgerald et al., 2009; Cho et al., 2016; Klejch et al., 2017), which aims to produce a formal textual report of a particular recording. In reality, speakers often repeat certain words multiple times, correct themselves in the middle of a sentence or use informal lan- guage to describe their ideas. In many applications such as meeting reports, automatic subtitling of presentations or subsequent machine translation of the ASR output, we would like the transcripts to be a formal-language equiva- lent of what was said and not the exact copy of the actually uttered words. In our paper, we present an extensive collection of Czech parliament plenary sessions which tries to contribute to both speech recognition per se as well as speech reconstruc- tion. The corpus consists of 444 hours of spoken Czech to- gether with the corresponding formal transcriptions as ob- tained by stenographers present at the meetings. Our paper is organized as follows. In Section 2., we de- scribe related literature and relevant work. In Section 3., we introduce our methodology used when constructing the corpus. Section 4. contains a detailed corpus description and its exploratory analysis. Section 5. presents the base- line results we obtained and we discuss them in …

Content truncated to 500 words