This paper describes a system for automated, quasi-real-time segmentation of short speech signals (e.g. single words) into phones. The phonetic representation of these signals must be given. If the orthographic representation is given an automatic rule-based grapheme-to-phoneme conversion is performed first. Rules are available for Dutch.

The segmentation algorithm proposed here attempts to determine phone boundaries by minimizing a cost function which takes into account the spectral variation of the signal as well as its resemblance with reference phones.

The algorithm is implemented as part of a program running on a TMS320-10/8086 biprocessor card. The program also offers real-time LPC analysis and synthesis.