Context-dependent grapheme-to-phoneme evaluation corpus using flexible contexts and Categorial Matrix

Abstract

Context-dependent pronunciation, e.g. homographs, is a difficult grapheme-to-phoneme conversion (G2P) issue. It causes accuracy downgrade in speech synthesis and speech recognition. However, the context-dependent pronunciation issue is rarely considered in collecting pronunciation corpus for evaluating accuracy of G2P. Thus, this paper proposes a context-dependent pronunciation corpus using grapheme-phoneme pairs with their context information for G2P assessment. The context information includes 1) Categorial Matrix for representing orthographic types and usage domains of orthographic groups (OG). Categorial Matrix is designed to investigate problem categories in the G2P. 2) regular-expression-based flexible context for representing context variation. 3) OG Classes for representing interchangeable OGs in the flexible context. The flexible context and the word classes are designed to remove redundant contexts while covering context variation with minimal sets of patterns. By using the proposed corpus, automatic context generation for G2P evaluation can be implemented.