Autolabelling Japanese TOBI.pdfVIP

AUTOLABELLING JAPANESE TOBI Nick Campbel l ATR Interpreting Telecommunications Research Lab oratories, 2-2 Hikaridai, Seika-cho, Kyoto 619-02, Japan. nick@itl.atr.co.jp ABSTRACT thesiser that predict the target pitch, p ower, and duration for text-to-sp eech and concept-to-sp eech pre-pro cessing. Focussing on the proso dic lab elling of Japanese as an exam- ple, this pap er describ es the application of sp eech synthesis 1.2. Aligning the corpus technology in a variety of sp eech pro cessing tasks. It dis- In order make use of the proso dic parameters extracted from cusses rst the use of synthesised utterances in the forced alignment and segmentation of a sp eech corpus, then the use a sp eech signal, it is necessary to asso ciate them with lin- guistic features of the sp eech by aligning the waveform with of generated proso dic contours to determine the proso dic phone lab els describing the content of the utterance. If this phrasing of an utterance, and nally the comparison with speech resynthesised using the proso dic transcription of the phone sequence is not available it must b e generated. The original utterance in order to check the transcription. It minimal requirements for pro cessing a sp eech corpus are closes with an analysis of results from an auto-transcription thus a) a representation of the sp eech waveform, and b) an of Japanese ToBI, and discusses some limitations of the pro- orthographic