7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Part-of-Speech Tagging in French Text-to-Speech Synthesis: Experiments in Tagset Selection

Hongyan Jing, Evelyne Tzoukermann

Lucent Technologies, USA

Part-of-speech tagging is needed for French Text-to-Speech (TTS) synthesis to disambiguate the pronunciation of homograph heterophones, liaison instances, and eventually to model intonational contours. A core problem in the part-of-speech tagging in French TTS is to decide on the tagset used for the tagger and the tagset needed by TTS. We carried out a number of experiments on several sizes of tagsets as well as on several algorithms to investigate this problem. Our experiment results suggest that there may be an optimal tagset to be used for the part-of-speech disambiguation in French TTS. This optimal tagset contains a slightly larger number of tags than the tagset that is needed by TTS for pronunciation disambiguation and intonational modeling purposes. In our experiments, the optimal tagset gives a 98.4% tagging accuracy for TTS, when a trigram Hidden Markov Model tagger is used.

Full Paper

Bibliographic reference.  Jing, Hongyan / Tzoukermann, Evelyne (2002): "Part-of-speech tagging in French text-to-speech synthesis: experiments in tagset selection", In ICSLP-2002, 97-100.