TTS pre-processing issues for mixed language support

Abstract

The design of an open domain Text-to-Speech (TTS) system mandates a preprocessing module that is responsible for preparing the input text so that it can be handled in a standard fashion by other TTS components. There are several pre-processing issues that are common across input domains, such as number, date and acronym handling. Others may be specific to the type of input under examination. In designing a TTS system supporting Maltese for SMS messages in the local context, specific issues are also encountered. In terms of language issues, the practical use of Maltese is heavily characterised by code-switching into other languages, most commonly in English. While the resulting language may not be considered ‘Maltese’ in the strictest sense of the language definition, it creates a state of affairs that cannot simply be ignored by a general purpose system. In respect of the SMS domain, the use of various shorthand notation and lack of phrase structure is encountered. This paper describes these specific issues in further detail and discusses techniques with which they may be addressed.