Development and Evaluation of Speech Synthesis Corpora for Latvian

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6633–6637 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6633 Development and Evaluation of Speech Synthesis Corpora for Latvian Roberts Darģis1, Pēteris Paikens1, Normunds Grūzītis1, Ilze Auziņa1, Agate Akmane2 1Institute of Mathematics and Computer Science, University of Latvia, 29 Raina Blvd., Riga, Latvia 2Riga East Clinical University Hospital, 2 Hipokrata Street, Riga, Latvia roberts.dargis@lumii.lv Abstract Text to speech (TTS) systems are necessary for any language to ensure accessibility and availability of digital language services. Recent advances in neural speech synthesis have enabled the development of such systems with a data-driven approach that does not require significant development of language-specific tools. However, smaller languages often lack speech corpora that would be sufficient for training current neural TTS models, which require at least 30 hours of good quality audio recordings from a single speaker in a noiseless environment with matching transcriptions. Making such a corpus manually can be cost prohibitive. This paper presents an unsupervised approach to obtain a suitable corpus from unannotated recordings using automated speech recognition for transcription, as well as automated speaker segmentation and identification. The proposed method and software tools are applied and evaluated on a case study for developing a corpus suitable for Latvian speech synthesis based on Latvian public radio archive data. Keywords: speech corpus development, text-to-speech, speech recognition, speaker segmentation 1. Introduction The popularity of speech synthesis as a topic in natural language processing has significantly increased after the publication of results by DeepMind (van den Oord et al., 2016), Baidu (Arik et al., 2017) and Google (Wang et al,, 2017; Shen et al., 2018), demonstrating the ability to create natural sounding speech with neural methods. While the new methods themselves are mostly language agnostic, extending these text to speech (TTS) systems for a particular language requires a specialized speech corpus for that language. Training these currently mainstream speech synthesis methods from scratch to adequate quality require about 30 hours of good quality audio recordings from a single speaker in noiseless environment, and an accurate transcription of these recordings. For less resourced languages, obtaining such a corpus is the major obstacle to development of TTS solutions. In this paper an unsupervised approach for such corpus creation is presented, using ASR and speaker segmentation as main components. 2. Related Work Before this research, a suitable corpus was not available for Latvian: there was a sizeable transcribed speech corpus (Pinnis et al., 2014) developed for automated speech recognition (ASR) purposes, but, because of ASR requirements, the corpus intentionally had a wide diversity of speakers, and did not have sufficient data from any single speaker to train a good TTS system. The usual approach to develop such a corpus involves narrators who record especially selected text segments in studio conditions. This approach has been used in multiple popular speech corpora of other languages, such as CMU ARTIC (Kominek et al., 2017) and CSTR VCTK Corpus (Veaux et al., 2017). The …

Content truncated to 500 words