Modified re-synthesis of initial voiceless plosives by concatenation of speech from different speakers

This paper describes a method of resynthesising utterance-initial voiceless plosives, given an original utterance by one speaker and a speech database of utterances by many other speakers. The system removes an initial voiceless plosive from an utterance and replaces it with another voiceless plosive selected from the speech database. (For example, if the original utterance was /tat/, the resynthesised utterance could be /k+at/.) In the method described, techniques used in general concatenative speech synthesis were applied in order to find those segments in the speech database that would yield the smoothest concatenation with the original segment. Results from a small listening test reveal that the concatenated samples are most often correctly identified, but that there is room for improvement on naturalness. Some routes to improvement are suggested.