Speaker-independent recognition of connected Swedish digits

This paper describes an approach towards speaker-independent recognition of connected Swedish digits. Each word is modelled as a sequence of acoustic-phonetic events that must be identified in order for the word to be recognized. The events are characterized by extremal values or rapid changes in formant, cepstrum, power and zero crossing contours. Preliminary results indicate a digit recognition accuracy of close to 80% when using digit strings of unknown length with 15-20 connected digits recorded through a telephone handset.