Building Open Javanese and Sundanese Corpora for Multilingual Text-to-Speech

Building Open Javanese and Sundanese Corpora for Multilingual Text-to-Speech Jaka Aris Eko Wibawa2, Supheakmungkol Sarin1, Chenfang Li3, Knot Pipatsrisawat1, Keshan Sodimana, Oddur Kjartansson1, Alexander Gutkin1, Martin Jansche1, Linne Ha1 1Google,2Google (on contract from Pactera Technologies NA Inc),3Google (on contract from Optimum Solutions Pte Ltd) {jwibawa,mungkol,chenfangl,thammaknot,ksodimana,oddur,agutkin,mjansche,linne}@google.com Abstract We present multi-speaker text-to-speech corpora for Javanese and Sundanese, the second and third largest languages of Indonesia spoken by well over a hundred million people. The key objectives were to collect high-quality data in an affordable way and to share the data publicly with the speech community. To achieve this, we collaborated with two local universities in Java and streamlined our record- ing and crowdsourcing processes to produce corpora consisting of 5,800 (Javanese) and 4,200 (Sundanese) mixed-gender recordings. We used these corpora to build several configurations of multi-speaker neural network-based text-to-speech systems for Javanese and Sundanese. Subjective evaluations performed on these configurations demonstrate that multilingual configurations for which Javanese and Sundanese are trained jointly with a larger corpus of Standard Indonesian significantly outperform the systems constructed from a single language. We hope that sharing these corpora publicly and presenting our multilingual approach to text-to-speech will help the community to scale up text-to-speech technologies to other lesser resourced languages of Indonesia. Keywords: low-resource languages, corpora, multilingual, text-to-speech 1. Introduction One of the important trends in modern speech and language technology is the increasing focus on scaling up the current state of the art to the large number of language commu- nities in the world. Progress in this direction is difficult, because, more often than not, the languages in the long tail of the distribution of the majority of the world’s languages lack adequate linguistic resources required for supporting speech and natural language research (Besacier et al., 2014; O’Horan et al., 2016). A prime example of such a challenge are the languages of Indonesia. Lewis et al. (2015) identify 707 distinct liv- ing languages (not dialects) that are spoken throughout the archipelago by over 255 million people. About twenty of these languages are spoken by over a million people. Standard Indonesian is the official national language of In- donesia and is spoken natively or as a second language by more than 200 million people (Paauw, 2009). Javanese with roughly 90 million native speakers and Sundanese with ap- proximately 40 million native speakers constitute the two largest regional languages of Indonesia. Unlike Indonesian, which received a lot of attention over the years, e.g. (Sakti et al., 2008; Manurung et al., 2010; Koto, 2016), both Ja- vanese and Sundanese are currently under-resourced due to the lack of openly available high-quality corpora. We address this problem for Javanese and Sundanese by building crowdsourced text-to-speech corpora for these lan- guages. This work is part of an ongoing effort by Google to build multi-speaker corpora for low-resource languages in an affordable way and to share the resulting data pub- licly. Building on our initial work on Bengali (Gutkin et al., 2016), Sinhala and other languages, we aimed to de- sign the …

Content truncated to 500 words