Fast Bootstrapping of Grapheme to Phoneme System for Under-resourced Languages - Application to the Iban Language

摘要

This paper deals with the fast bootstrapping of Grapheme-to-Phoneme (G2P) conversion system, which is a key module for both automatic speech recognition (ASR), and text-to-speech synthesis (TTS). The idea is to exploit language contact between a local dominant language (Malay) and a very under-resourced language (Iban - spoken in Sarawak and in several parts of the Borneo Island) for which no resource nor knowledge is really available. More precisely, a pre-existing Malay G2P is used to produce phoneme-sequences of Iban words. The phonemes are then manually post-edited (corrected) by an Iban native. This resource, which has been produced in a semi-supervised fashion, is later used to train the first G2P system for Iban language. As a by-product of this methodology, the analysis of the "pronunciation distance" between Malay and Iban enlighten the phonological and orthographic relations between these two languages. The experiments conducted show that a rather efficient Iban G2P system can be obtained after only two hours of post-edition (correction) of the output of Malay G2P applied to Iban words.