We propose an original knowledge based approach of acoustic-phonetic decoding of continuous speech relying on the use of triplets: a phone with its phonetic context. A triplet is made up of two description levels: an acoustic description in terms of acoustic 'events (formant, burst ...) and an expert component which indicates the significant acoustic correlates to recognize a triplet. The matching process is firstly performed at the acoustic level, results are then weighted with the help of the expert component. As triplet are phone prototypes it is possible to extract acoustic relations between two reference triplets. These relations must be simultaneously satisfied between two triplet instances of the unknown sentence and two reference triplets proposed as solution. Relaxation techniques enable to implement this idea in order to increase consistency of the global solution.