Speech Rate Calculations with Short Utterances: A Study from a Speech-to-Speech, Machine Translation Mediated Map Task

Speech Rate Calculations with Short Utterances: A Study from a Speech-to-Speech, Machine Translation Mediated Map Task Hayakawa Akira, 1 Carl Vogel, 1 Saturnino Luz, 2 Nick Campbell 1 1 School of Computer Science and Statistics, Trinity College Dublin, Ireland 2 Usher Institute of Population Health Sciences & Informatics, University of Edinburgh, UK campbeak@tcd.ie, vogel@cs.tcd.ie, S.Luz@ed.ac.uk, nick@tcd.ie Abstract The motivation for this paper is to present a way to verify if an utterance within a corpus is pronounced at a fast or slow pace. An alternative method to the well-known Word-Per-Minute (wpm) method for cases where this approach is not applicable. For long segmentations, such as the full introduction section of a speech or presentation, the measurement of wpm is a viable option. For short comparisons of the same single word or multiple syllables, Syllables-Per-Second (sps) is also a viable option. However, when there are multiple short utterances that are frequent in task oriented dialogues or natural free flowing conversation, such as those of the direct Human-to-Human dialogues of the HCRC Map Task corpus or the computer mediated inter-lingual dialogues of the ILMT-s2s corpus, it becomes difficult to obtain a meaningful value for the utterance speech rate. In this paper we explain the method used to provide a alternative speech rate value to the utterance of the ILMT-s2s corpus and the HCRC Map Task corpus. Keywords: speech rate, utterance duration comparison, task oriented dialogues 1. Introduction Computer mediated communication is becoming more fre- quent. The next step in this new communication style, is Inter-Lingual Computer Mediated Communication. Re- cently, Microsoft released the Skype Translator (Lewis, 2015) that translates up to 10 languages in Speech-to- Speech Machine Translation. The Japanese Ministry of Internal Affairs and Communication has announced that the Tokyo Olympics in 2020 are to employ information systems that use multilingual machine mediated commu- nication in Speech-to-Speech (S2S) form for 17 languages, using the VoiceTra system (Matsuda et al., 2013). Com- puter mediated multi-lingual communication will create a conversation style that few users have adequate exposure (Hara and Iqbal, 2015) and may have difficulty adapting to. For example, using the speech rate calculation method ex- plained in this paper, gender differences in the adaptation speech rate have been identified (Hayakawa et al., 2017b) but with little to no improvement to the automatic speech recognition results. With users of computer mediated in- teraction, the “Speakers are unlikely to have a good model of what computers are likely to know, presumably in part because natural language systems tend to be tailored to par- ticular and quite specific uses.” (Branigan et al., 2010, p. 2366). Being able to provide feedback to the user about the changes to their speech rate, should help prevent the hyper- articulation reported by Hayakawa et al. (2017b). The method in this paper is not new in theory. Sztah´o et al. (2015) use a similar method of comparing the syllable duration it takes a subject to utter a specific phrase with pre- vious recordings of the utterance by the same …

Content truncated to 500 words