Crowdsourcing Latin American Spanish for Low-Resource Text-to-Speech

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6504–6513 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6504 Crowdsourcing Latin American Spanish for Low-Resource Text-to-Speech Adriana Guevara-Rukoz1,2,3, I¸sın Demir¸sahin1, Fei He1, Shan-Hui Cathy Chu1, Supheakmungkol Sarin1, Knot Pipatsrisawat1, Alexander Gutkin1, Alena Butryna1, Oddur Kjartansson1 1Google Research, Japan, Singapore, United States and United Kingdom 2Laboratoire de Sciences Cognitives et Psycholinguistique, ENS/EHESS/CNRS/PSL/INRIA, France 3Graduate School of Engineering, The University of Tokyo, Japan {aguekoz,isin,mungkol,thammaknot,agutkin}@google.com Abstract In this paper we present a multidialectal corpus approach for building a text-to-speech voice for a new dialect in a language with existing resources, focusing on various Latin American dialects of Spanish. We first present public speech datasets for Argentinian, Chilean, Colombian, Peruvian, Puerto Rican, and Venezuelan Spanish, specifically constructed with text-to-speech applications in mind using crowdsourcing. We then compare monodialectal TTS voices built with minimal data to voices made with a multidialectal model built by pooling all the resources from all dialects. Our results show that the multidialectal model outperforms the monodialectal baseline models. We also experiment with a “zero-resource” dialect scenario where we build a multidialectal voice for a dialect while holding out target dialect recordings from the training data. Keywords: corpora, text-to-speech, phonology, Spanish dialects, Latin America 1. Introduction Building a high-quality text-to-speech (TTS) voice is costly and requires significant effort both in terms of collecting nontrivial amounts of recorded speech in each language, as well as building the corresponding linguistic resources. In low-resource language scenarios several approaches have been proposed to address the issue of data and resource scarcity. These techniques range from utilizing general speech data found online that was not necessarily recorded for speech applications in mind (Baljekar, 2018; Cooper, 2019) to pooling the data from multilingual corpora (Li and Zen, 2016). It has been shown that constructing the mod- els from multiple speakers (Gutkin et al., 2016), as well as using the data from related languages (i.e., languages from close families or language areas (Emeneau, 1956)), consis- tently improves the quality of the voice with minimal re- sources (Baljekar et al., 2018; Wibawa et al., 2018; Demir- sahin et al., 2018). We believe that this approach can also be extended to constructing the models for new dialects of a given language. In this paper we explore different compositions of datasets for building voices for Argentinian (AR), Chilean (CL), Pe- ruvian (PE) and Venezuelan (VE) Spanish dialects using multi-speaker data crowdsourced1 in each locale, combin- ing it with the professional studio recordings of Peninsu- lar (ES) and United States (US)2 Spanish dialects. Past TTS research on Argentinian (Torres et al., 2012; Violante, 2012), Peruvian (Florentino, 2016), Venezuelan (Rodríguez et al., 2006) and other Latin American Spanish dialects, such as Colombian (Correa et al., 2010), focused mostly on single-speaker nonparametric concatenative systems (Hunt and Black, 1996). An arguably less brittle, parametric, ap- proach to TTS (Zen et al., 2013; Zen et al., 2016), mediated 1As opposed to data recorded by professional voice actors. …

Content truncated to 500 words