The SI TEDx-UM speech database: a new Slovenian Spoken Language Resource

The SI TEDx-UM Speech Database: a new Slovenian Spoken Language Resource Andrej Žgank, Mirjam Sepesy Maučec, Darinka Verdonik University of Maribor, Faculty of Electrical Engineering and Computer Science, Institute of Electronics and Telecommunications, Maribor, Slovenia E-mail: andrej.zgank@um.si Abstract This paper presents a new Slovenian spoken language resource built from TEDx Talks. The speech database contains 242 talks in total duration of 54 hours. The annotation and transcription of acquired spoken material was generated automatically, applying acoustic segmentation and automatic speech recognition. The development and evaluation subset was also manually transcribed using the guidelines specified for the Slovenian GOS corpus. The manual transcriptions were used to evaluate the quality of unsupervised transcriptions. The average word error rate for the SI TEDx-UM evaluation subset was 50.7%, with out of vocabulary rate of 24% and language model perplexity of 390. The unsupervised transcriptions contain 372k tokens, where 32k of them were different. Keywords: spoken language resource, under-resourced language, unsupervised transcriptions 1. Introduction Under-resourced languages still present a great challenge for the speech processing community. Slovenian belongs to such a group of under-resourced languages and is, with 2 million speakers, one of the smallest official EU languages. The development of Slovenian speech technology systems started in the end of 80’s. Special emphasis was given to the development of spoken language resources. During these years, comparable categories of language resources to the main ones for English were built: SNABI and GOPOLIS, (TIMIT, Resource Management like), 1000 FDB SpeechDat(II) and PoliDat (SpeechDat like), BNSI Broadcast News and SiBN Broadcast News. The common characteristics of these language resources are that they were manually annotated and transcribed. This presents a time consuming and expensive task, especially for under-resourced languages like Slovenian. The first Slovenian language resource that differs from this approach was SloParl with parliamentary debates, which was based on imperfect transcriptions generated in parliament. The size of these Slovenian language resources is approximately 200 hours of speech, which is significantly less than for frequently spoken world languages. Additional drawback for developing speech recognition systems is the fact that Slovenian language belongs to the group of highly-inflected languages. This group of languages usually needs significantly more spoken training material to successfully develop high quality large vocabulary continuous speech recognition systems. In order to increase the amount of available language resources for Slovenian automatic speech recognition and other speech technology systems, the decision was made to build a new Slovenian resource based on TEDx Talks (Technology, Entertainment, Design) using automatic acoustic classification and large vocabulary continuous speech recognition. The costs and time usually needed to build such a speech resource were reduced with annotating and transcribing the acquired speech mainly in an unsupervised way. The paper is organized as follows. The Section 2 presents the acquisition process. The automatic segmentation and transcription are described in Section 3. The manual annotation and transcription of development and evaluation set are presented in Section 4. The finalized speech database is described in Section 5, while the conclusion is given in Section 6. …

Content truncated to 500 words