Phonetic decoding of continuous speech with the APHODEX expert system

In order to increase the accuracy of continuous speech acousticphonetic decoding, we started the APHODEX project some years ago. Our aim is to develop an expert system that implements the knowledge of an expert spectrogram reader, the phonetician F. Lonchamp. In the present version of the system, procedural and declarative approaches have been used jointly for the representation of phonetic expertise relating to : the segmentation of speech into phonemelike units, the classification of segments into phonetic categories (coarse !abeling) and subsequent interpretation in terms of phones (fine labeling). APHODEX is capable of processing hypotheses in parallel, using contextual analysis, fuzzy and uucertain reasoning, especially for the interpretation of cues and featureH extracted from the speech signal. Decoding results are represented in the form of a phonelattice; the description of alternative segmentations and labelings is possible. A prototype version of APHODEX has been implemented, using the SNORRl speech processing environment. Results concerning a multi-speaker corpus of continuous speech are presented and discussed.