Customizing Grapheme-to-Phoneme System for Non-Trivial Transcription Problems in Bangla Language

Proceedings of NAACL-HLT 2019, pages 3191–3200 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3191 Customizing Grapheme-to-Phoneme System for Non-Trivial Transcription Problems in Bangla Language Sudipta Saha Shubha1, Nafis Sadeq1, Shafayat Ahmed1, Md. Nahidul Islam1, Muhammad Abdullah Adnan1, Md. Yasin Ali Khan2, and Mohammad Zuberul Islam2 1Bangladesh University of Engineering and Technology (BUET) 2Samsung R&D Institute, Bangladesh {sudipta,nafis,shafayat,nahid.rimon}@ra.cse.buet.ac.bd adnan@cse.buet.ac.bd,{yasin.ali,m.zuberul}@samsung.com Abstract Grapheme to phoneme (G2P) conversion is an integral part in various text and speech process- ing systems, such as: Text to Speech system, Speech Recognition system, etc. The existing methodologies for G2P conversion in Bangla language are mostly rule-based. However, data-driven approaches have proved their su- periority over rule-based approaches for large- scale G2P conversion in other languages, such as: English, German, etc. As the performance of data-driven approaches for G2P conversion depend largely on pronunciation lexicon on which the system is trained, in this paper, we investigate on developing an improved train- ing lexicon by identifying and categorizing the critical cases in Bangla language and include those critical cases in training lexicon for de- veloping a robust G2P conversion system in Bangla language. Additionally, we have incor- porated nasal vowels in our proposed phoneme list. Our methodology outperforms other state- of-the-art approaches for G2P conversion in Bangla language. 1 Introduction Grapheme to phoneme (G2P) conversion provides a mapping between a word and its pronunciation. Such mapping provides opportunity for a non- native person to learn the correct pronunciation of words of a foreign language. Moreover, in mod- ern Text to Speech (TTS) and Automatic Speech Recognition (ASR) systems, G2P conversion is an integral task. The task of G2P conversion is generally lan- guage specific due to language specific conven- tions, rules, pronunciation constraints, etc. In this paper, we focus on Modern Standard Bangla. An example of G2P conversion in Bangla language: phonetic transcription of অনুশীলন (practice) is /o n u sh i l O n/. (Please refer to Table 1 for our phoneme symbols.) The simplest means of G2P conversion is to build up a lexicon or dictionary containing the mapping from words to their corresponding pro- nunciations. However, it fails to provide pronun- ciations for unknown words and inclusion of newer words increases memory requirement. In another approach (Mosaddeque et al., 2006), there are pre- defined rules for the conversion of a word to its pronunciation. Though such rule-based approach can work for any word, the system becomes com- plex when it tries to formulate rules for incorpo- rating all irregularities of pronunciation in a lan- guage. Clearly, these approaches are not feasible for large-scale G2P conversion which is necessary in any modern TTS or ASR system. Data-driven ma- chine learning approaches have great potential in such large-scale G2P conversion (Rao et al., 2015). In such an approach, a machine learning model predicts the phoneme conversion of a grapheme, being trained on a lexicon. A predominant work following such approach in Bangla language is by Google (Gutkin et al., 2016), where they train …

Content truncated to 500 words