Optimization of letter to sound rules construction
Matus Vasek, Gregor Rozinaj · 2013
This article falls within the field of speech processing - speech synthesis. The content is focused on phonetic transcription of the text. The transcript rules as well as dictionary of phonetic transcriptions are important tools for the phonetic transcription. The article deals with the transcript rules called LTS (Letter-To-Sound). Automatically generated LTS rules are presented in form of binary decision trees as a solution especially suitable for flexible languages. This article introduces an idea how to improve and optimize the training process of LTS rules (data-driven approach). Optimization consists of two phases. The first phase lies in the training data preprocessing and in the training database preparation. The second phase aims to achieve an optimal size of LTS rules regarding the size of phonetic transcription dictionary. This paper introduces a method how to generate residual dictionary of exceptions. The size of generated LTS rules has a direct impact on the size of residual (delta) dictionary. The paper describes the comparison between individual iterations of LTS rules training process. Iterations differ from each other due to a different threshold parameter. Finally article presents the results achieved using this method to select the most optimal threshold parameter.