CPJD Corpus: Crowdsourced Parallel Speech Corpus of Japanese Dialects

CPJD Corpus: Crowdsourced Parallel Speech Corpus of Japanese Dialects Shinnosuke Takamichi, Hiroshi Saruwatari Graduate School of Information Science and Technology, The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan {shinnosuke takamichi, hiroshi saruwatari}@ipc.i.u-tokyo.ac.jp Abstract Public parallel corpora of dialects can accelerate related studies such as spoken language processing. Various corpora have been collected using a well-equipped recording environment, such as voice recording in an anechoic room. However, due to geographical and expense issues, it is impossible to use such a perfect recording environment for collecting all existing dialects. To address this problem, we used web-based recording and crowdsourcing platforms to construct a crowdsourced parallel speech corpus of Japanese dialects (CPJD corpus) including parallel text and speech data of 21 Japanese dialects. We recruited native dialect speakers on the crowdsourcing platform, and the hired speakers recorded their dialect speech using their personal computer or smartphone in their homes. This paper shows the results of the data collection and analyzes the audio data in terms of the signal-to-noise ratio and mispronunciations. Keywords: crowdsourcing, Japanese dialect, speech, transcription 1. Introduction A public corpus of low-resourced languages can accelerate related studies, such as natural language, spoken language, and speech signal processing. Thanks to recent improve- ments in machine learning techniques for these research ar- eas (v. d. Oord et al., 2016; Hinton et al., 2012; Takamichi et al., 2017), the accuracies in the rich-resourced languages (such as English, Chinese, and Japanese) have become high, and attention is shifting to more challenging tasks, including language and speech processing of dialects. Di- alect processing is one of the topics being actively targeted in machine translation (Salloum et al., 2014), speech recog- nition (Hirayama et al., 2014), speech perception (Jacewicz and Fox, 2017), and speech synthesis (Masmoudi et al., 2016). For studies on Japanese dialect, (Yoshino et al., 2016) collected a parallel database of Japanese dialects with voice recording in their well-equipped recording environ- ment. However, geographical and expense issues make it impossible to use such a perfect recording environment for all existing Japanese dialects. In this work, we collected parallel data of the common lan- guage and several dialects in Japanese using a web-based recording and crowdsourcing platform. We recruited native dialect speakers on the crowdsourcing platform, and the hired speakers of 21 dialects uttered 250 sentences for each. They also converted text in the common language into their dialect, and read the dialect text in the web-based recording platform available on personal computers and smartphones. Thus, the resultant resources consist of texts (one common language and 21 dialects), nine hours of speech data, and their geographic contexts (i.e., areas of the dialect). We an- alyzed the collected speech data from perspectives of the speech signal and spoken language processing. This paper shows the signal-to-noise ratio (SNR) results for the record- ing environments and the number of mispronunciations. 2. Existing resources of Japanese dialects and their problems Several text/speech corpora of Japanese dialects have been compiled (Kubozomo, 2001 2008; National Institute for Japanese Language and …

Content truncated to 500 words