DaCToR: A Data Collection Tool for the RELATER Project

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6627–6632 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6627 DaCToR: A Data Collection Tool for the RELATER Project †Juan Hussain, +†Oussama Zenkri, †Sebastian St¨uker, ∗†Alexander Waibel †Institute for Anthropomatics and Robotics KIT - Karlsruhe Institute of Technology, Germany firstname.lastname@kit.edu + firstname.lastname@student.kit.edu ∗ Language Technologies Institute Carnegie Mellon University firstname.lastname@cmu.edu Abstract Collecting domain-specific data for under-resourced languages, e.g., dialects of languages, can be very expensive, potentially financially prohibitive and taking long time. Moreover, in the case of rarely written languages, the normalization of non-canonical transcription might be another time consuming but necessary task. In order to collect domain-specific data in such circumstances in a time and cost-efficient way, collecting read data of pre-prepared texts is often a viable option. In order to collect data in the domain of psychiatric diagnosis in Arabic dialects for the project RELATER, we have prepared the data collection tool DaCToR for collecting read texts by speakers in the respective countries and districts in which the dialects are spoken. In this paper we describe our tool, its purpose within the project RELATER and the dialects which we have started to collect with the tool. Keywords: speech, data collection tool, under-resourced languages, dialects 1. Introduction Speech recognition, speech translation and speech synthe- sis systems have a very wide range of applications today. These systems, which employ modern learning algorithms usually require large amount of data to achieve acceptable performance. However, Obtaining sufficient amount of data is often a problem for low-resource languages, such as di- alects or non-written languages, and for domain-specific fields, such as the medical domain. For training speech recognition systems one needs to collect sufficient amounts of spoken data that is transcribed at sentence level. Collect- ing data in the exact domain to be addressed in the required speaking style, such as spontaneous speech in psychiatric interviews, is often infeasible. Instead, voice data is often collected from the Internet and transcribed manually (Her- nandez et al., 2018). Another common source of voice data may be telephone calls (Canavan, Alexandra, David Graff, and George Zipperlen, 1997) or news broadcasts (Walker, Kevin, et al., 2014). These methods require significant transcription effort, especially if no automatic transcription system is available. In the case of rare dialects, another problem can be the ambiguity of the transcription created by different transcribers, which requires additional post- processing effort for normalization. In order to collect data in a cost efficient way and avoid the transcription effort, we record domain-specific texts that are read by appropri- ate speakers of the language or dialect in question. Con- sequently, we avoid the problems of non-canonical tran- scriptions and writing systems. Though read speech is not necessarily the required speaking style, it is possible in this way to collect a variety of speech data stemming from many speakers. This contributes to improved speaker and chan- nel independence in case of multiple recordings of the same text. As …

Content truncated to 500 words