New Grapheme Generation Rules for Two-Stage Modelbased Grapheme-to-Phoneme Conversion

Abstract

The precise conversion of arbitrary text into its corresponding phoneme sequence (grapheme-to-phoneme or G2P conversion) is implemented in speech synthesis and recognition, pronunciation learning software, spoken term detection and spoken document retrieval systems. Because the quality of this module plays an important role in the performance of such systems and many problems regarding G2P conversion have been reported, we propose a novel two-stage model-based approach, which is implemented using an existing weighted finite-state transducer-based G2P conversion framework, to improve the performance of the G2P conversion model. The first-stage model is built for automatic conversion of words to phonemes, while the second-stage model utilizes the input graphemes and output phonemes obtained from the first stage to determine the best final output phoneme sequence. Additionally, we designed new grapheme generation rules, which enable extra detail for the vowel and consonant graphemes appearing within a word. When compared with previous approaches, the evaluation results indicate that our approach using rules focusing on the vowel graphemes slightly improved the accuracy of the out-of-vocabulary dataset and consistently increased the accuracy of the in-vocabulary dataset.

Using Reversed Sequences and Grapheme Generation Rules to Extend the Feasibility of a Phoneme Transition Network-Based Grapheme-to-Phoneme Conversion

“…Over the last few years, it has proven considerably difficult to improve the performance of a G2P conversion model for OOV words, because each method or approach is uniquely designed using different techniques to address particular challenges. It is seemingly impossible to utilize any single method to deal comprehensively with the host of problems encountered by G2P conversion [12]. Therefore, we designed a PTN-based architecture for G2P conversion that allows many different methods to be applied together, in order to deal broadly with the various problems.…”

“…Even if the reversed g-p sequences can make a complementary model that can generate an additional phonemesequence hypothesis for each source method, the risk from combining different methods nevertheless remains. Hence, we introduce a novel use of grapheme generation rules (GGRs) [12] to minimize this risk. This allows us to use only a single method for implementing a PTN-based G2P conversion model.…”

“…In orthographically complex languages such as English, the interaction between vowels in a word significantly affects the spelling. Hence, a technique for generating new grapheme sequences from the same input text (known as GGRs) has been proposed for adding extra-sensitive information to each vowel-grapheme appearing in a word [12]. Suppose that a grapheme sequence g = g 1 g 2 .…”

“…However, each of these methods has been designed using specific techniques that address particular challenges faced by G2P conversion. Therefore, any single approach will not suffice when addressing all of the problems encountered by G2P conversion [12]. Considering this, a combination of various approaches using different methods is a reasonable strategy for treating these problems in a flexible manner.…”

“…In order to mitigate this risk, we selected a minimum number of methods. † We also present a novel use for rightto-left (reversed) grapheme-phoneme (g-p) sequences and grapheme generation rules (GGRs) [12]. In this study, both techniques are especially helpful for extending the feasibility and improving the performance of PTN-based G2P conversion, because they increase the number of phonemesequence hypotheses without increasing the number of methods used.…”