Burmese Speech Corpus, Finite-State Text Normalization and Pronunciation Grammars with an Application to Text-to-Speech

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6328–6339 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6328 Burmese Speech Corpus, Finite­State Text Normalization and Pronunciation Grammars with an Application to Text­to­Speech Yin May Oo†, Theeraphol Wattanavekin, Chenfang Li†, Pasindu De Silva, Supheakmungkol Sarin, Knot Pipatsrisawat, Martin Jansche†, Oddur Kjartansson, Alexander Gutkin Google Research Singapore, United States and United Kingdom {mungkol,thammaknot,oddur,agutkin}@google.com Abstract This paper introduces an open­source crowd­sourced multi­speaker speech corpus along with the comprehensive set of finite­state transducer (FST) grammars for performing text normalization for the Burmese (Myanmar) language. We also introduce the open­source finite­state grammars for performing grapheme­to­phoneme (G2P) conversion for Burmese. These three components are necessary (but not sufficient) for building a high­quality text­to­speech (TTS) system for Burmese, a tonal Southeast Asian language from the Sino­Tibetan family which presents several linguistic challenges. We describe the corpus acquisition process and provide the details of our finite state­based approach to Burmese text normalization and G2P. Our experiments involve building a multi­speaker TTS system based on long short term memory (LSTM) recurrent neural network (RNN) models, which were previously shown to perform well for other languages in a low­resource setting. Our results indicate that the data and grammars that we are announcing are sufficient to build reasonably high­quality models comparable to other systems. We hope these resources will facilitate speech and language research on the Burmese language, which is considered by many to be low­resource due to the limited availability of free linguistic data. Keywords: speech corpora, finite­state grammars, low­resource, text­to­speech, open­source, Burmese 1. Introduction The Burmese (or Myanmar) language belongs to the Sino­ Tibetan language family. It is the largest non­Chinese lan­ guage from that family (Jenny and Tun, 2016) with the to­ tal number of speakers in Myanmar and abroad is estimated at around 43 million: 33 million native (L1) speakers with the additional population of 10 million second­language (L2) speakers (SIL International, 2019). Due to the rela­ tive scarcity of the available language resources Burmese is often considered to be an under­resourced language (Gold­ hahn et al., 2016). With regard to speech resources, and in particular to text­ to­speech (TTS), this scarcity is especially pronounced. Unlike the automatic speech recognition (ASR), TTS cor­ pora for state­of­the­art or commercial systems has tra­ ditionally been recorded in professional studios by dedi­ cated voice actors. This process is both expensive (good voices are hard to find) and time consuming (consider­ able effort is spent on supervising the recordings and main­ taining a steady high audio quality). With the advance of new approaches that range from utilizing the found data for under­resourced languages (Baljekar, 2018; Cooper, 2019), crowd­sourcing (Gutkin et al., 2016) to multi­speaker and multilingual sharing (Li and Zen, 2016; Chen et al., 2019; Nachmani and Wolf, 2019), building TTS systems for under­resourced languages has become simpler (Wibawa et al., 2018; Prakash et al., 2019). However, the availability of a Burmese TTS corpora that is free for all is still …

Content truncated to 500 words