[Segmentation and phonetic classification by the DAWID system (author's transl)]

The DAWID speech recognition system, which is briefly described in this paper, performs an acoustical analysis which in some respect can be compared with a phonetic transcription. The limitations of this comparison are discussed. The segmentation and phonetic classification performance of the system is investigated on the basis of 5,520 spoken words (46 different one- and two-syllable words, 12 male speakers). The system performance is examined with the aid of varying classes of speech sounds embedded in different phonetic context. The results indicate that the DAWID system reliably transforms some classes of speech sounds into sequences of symbols, the elements of which represent 'quasi-allophonic' sound classes.