SEGRE : AN AUTOMATIC TOOL FOR GRAPHEME-TO-ALLOPHONE TRANSCRIPTION IN CATALAN

Segre is a rule-based automatic phonetic transcription system for Catalan, jointly developed by the Universitat Politècnica de Catalunya, the Universitat Autònoma de Barcelona and the Universitat de Barcelona in the framework of the Catalan Reference Centre for Language Engineering (CREL, Centre de Referència en Enginyeria Lingüística). The syntax of the rules has been designed to obtain phonetic transcriptions for four major dialects of Catalan: the Central or Eastern dialect, spoken in the East of Catalonia, the North-Western or Western dialect, spoken in the West of Catalonia (including the South), the Balearic, spoken in the Balearic Islands, and finally the Valencian, spoken in the Valencian Community. The transcriber has been designed in a very flexible way, since the rules are fed to the program, which has very little hardwired knowledge. They are defined externally and specified in a number of ASCII text files following a simple syntax for grapheme-to-allophone and allophone-to-allophone conversion rules, the latter necessary to obtain those modifications due to coarticulation phenomena across word boundaries. So, the tool provides a phonetic transcription, as broad or narrow as desired, in isolated mode (without coarticulation across word boundaries) or in text mode (with coarticulation between words). Furthermore, and due to the fact that the rules may be tweaked in any desired way, it is also possible, for instance, to transcribe particular subdialects or to obtain more or less narrow transcriptions. This follows from the fact that there is not a closed list of allophones in terms of which words are transcribed. Instead, the allophones are given by the various rule files. The accuracy of transcriptions of new texts, when compared with human expert generated transcriptions, is of 99.1% for isolated words and 99,39% for running text. Segre can be then considered a useful tool to model how coarticulation modifies the isolated transcription of words in real sentences. So, it is helpful not only to build speech synthesis systems but also to train subword-based speech recognition systems. Certainly, although in simple tasks such as connected digits or phonetic recognition no phonetic dictionary is needed, if the coarticulation rules are incorporated into the recognition network, they may complement the work done by cross-word units.