Parallel Speech Corpora of Japanese Dialects

Parallel Speech Corpora of Japanese Dialects Koichiro Yoshino1, Naoki Hirayama2,†, Shinsuke Mori3, Fumihiko Takahashi4,†, Katsutoshi Itoyama5, and Hiroshi G. Okuno5,6 1Graduate School of Information Science, Nara Institute of Science and Technology, Ikoma, 630-0192, Japan 2Industrial ICT Solutions Company, Toshiba Corporation, 3-22, Katamachi, Fuchu, 183-8512, Japan 3Academic Center for Computing and Media Studies, Kyoto University, Sakyo, Kyoto, 606-8501, Japan 4Yahoo Japan Corporation, Midtown Tower, 9-7-1 Akasaka, Minato-ku, Tokyo, 107-6211, Japan 5Graduate School of Informatics, Kyoto University, Sakyo, Kyoto, 606-8501, Japan 6School of Creative Science and Engineering, Waseda University, Ohkubo, Shinjyuku, Tokyo 169-0072, Japan koichiro@is.naist.jp, naoki2.hirayama@toshiba.co.jp, forest@i.kyoto-u.ac.jp, ftakahas@yahoo-corp.jp, itoyama@kuis.kyoto-u.ac.jp, okuno@aoni.waseda.jp Abstract Clean speech data is necessary for spoken language processing, however, there is no public Japanese dialect corpus collected for speech processing. Parallel speech corpora of dialect are also important because real dialect affects each other, however, the existing data only includes noisy speech data of dialects and their translation in common language. In this paper, we collected parallel speech corpora of Japanese dialect, 100 read speeches utterance of 25 dialect speakers and their transcriptions of phoneme. We recorded speeches of 5 common language speakers and 20 dialect speakers from 4 areas, 5 speakers from 1 area, respectively. Each dialect speaker converted the same common language texts to their dialect and read them. Speeches are recorded with closed-talk microphone, using for spoken language processing (recognition, synthesis, pronounce estimation). In the experiments, accuracies of automatic speech recognition (ASR) and Kana Kanji conversion (KKC) system are improved by adapting the system with the data. Keywords: Speech, Transcription, Dialect, Japanese 1. Introduction Texts and their read speeches are necessary for natural and spoken language processing (NLP and SLP). Tireless ef- forts of data collection of speech data and their transcrip- tions from the early stage of NLP and SLP researches dras- tically improved accuracies of a variety of NLP and SLP tasks. However, we still do not have enough resources for some minor languages, and it causes less accuracies in mi- nor languages (Kominek and Black, 2006). Especially, we do not have enough speech data of dialects. The lack of dialect speech data disturbs the use of NLP and SLP applications of dialect speakers. For example, it is very difficult to recognize dialect speeches accurately by using a model trained in common language. Automatic speech recognition (ASR) system is now generally used in pub- lic services, for example, taking minutes of national con- gresses and city councils (Akita et al., 2009). Such appli- cations are expected to be used for not only current for- mal situations but also more casual situations, for example, taking minutes of courts, and dialect speech recognition is necessary for this purpose. If we have a small parallel set of dialect and common lan- guage, SLP technologies developed in common language can be applied to the dialect rapidly. Japanese dialect speeches are collected and transcribed in linguistics area (National Institute for Japanese Language and Linguistics, 2001 2008), however, the sound data of the corpus is not collected for using on speech …

Content truncated to 500 words