Building of children speech corpus for improving automatic subtitling services

Building of children speech corpus for improving automatic subtitling services Matus Pleva, Stanislav Ondas, Daniel Hládek, Jozef Juhar, Ján Staš Department of Electronics and Multimedia Communications Technical University of Kosice, Slovakia matus.pleva@tuke.sk stanislav.ondas@tuke.sk daniel.hladek@tuke.sk jozef.juhar@tuke.sk jan.stas@tuke.sk Yuan-Fu Liao Department of Electronic Engineering National Taipei University of Technology yfliao@mail.ntut.edu.tw Abstract This paper describes the development and first evaluation of the new Slovak children speech audio corpus for improving the automatic broadcast news subtitling engine developed on the Technical University of Kosice in cooperation with the Slovak Academy of Sciences. The current automatic speech recognition (ASR) systems are reliable for a clean, prepared speech of adults with not very long pause inside the sentences. For speech recognition of children's, it is still a challenge from different reasons. They use much slang, and diminutive words, undeveloped pronunciation, shorter vocal tract (different speech parameters), the sentence syntax is different. The paper presents the results of the children speech automatic recognition from the system built for broadcast news transcription. Keywords: automatic speech recognition, children speech, audio corpus, annotation, database design. 1. Introduction The speech technology has significant potential, currently it has growing interest among children and technically enthusiastic people [1]. The International Speech and Communication Association (ISCA) has a Special Interest Group (SIG) for Child Computer Interaction (CHILD) [2] and is organizing a special Workshop on Child Computer Interaction - WOCCI and last years also Language Teaching, Learning and Technology - LTLT. This year a special The 2019 Conference on Computational Linguistics and Speech Processing ROCLING 2019, pp. 325-333 ©The Association for Computational Linguistics and Chinese Language Processing 325 session on the prestigious Interspeech conference will be held in September 2019 in Graz called Spoken Language Processing for Children's Speech [3]. The development of children's speech corpora for different languages is in progress [4] (British English, German and Swedish), [5] (non-native English), [6] Chinese, [7] Cantonese, [8] Jamaican English, [9] interactive emotional children speech and many others. For European union also small European languages are essential for electronic communication, so we decided to start the building of Slovak children speech corpus for improvement of the Slovak automatic speech recognition engines already built [10, 11]. Of course, the speech parameters different for children speech because of different vocal tract sizes [12], and they are many algorithms (Vocal-Tract Length Normalization - VTLN) how to handle it [13]. For children's speech, the formant frequencies are higher, the speech rate is slower or higher than in adult speech, and the language contains more home slang, garbled and imaginary words. The Slovak language belongs to a group of Slavic languages, which are typical of inflection and free word order, which means it is morphologically rich and uses a very large vocabulary [10, 14]. These features make the Slovak automatic speech recognition task very complicated, and a large amount of data is required for automatic large vocabulary spontaneous speech recognition [14]. This article describes the first step, the collection of the first data, manual annotation, and testing of the current ASR …

Content truncated to 500 words