Performance Evaluation Test of an Automatic Segmentation System for Italian and American-english Continuous Speech.

We present in this paper an automatic system for the segmentation of continuos speech. The segmentation algorithm has been implemented modifying the CDHMM system RES (Recognition Experimental System) for the automatic recognition of continuos speech. The segmentation task is performed as a phonetic recognition constrained by the grammar model built only on the exact sequence of uttered phonemes. We also report the results of some performance evaluation test using TIMIT (American-English) and APASCI (Italian) database. In order to evaluate the system performance also on telephonic application, we modified the Italian database to produce a telephonic version. Therefore we report both the results for normal and telephonic speech focusing our attention on the performance of vowel segmentation. As our envisaged application is the segmentation of speech recorded by means of wire-tapping for forensic application, we also report some specific results for the identification of phonemic boundaries on a real wire tapped speech.