The SPIN speaker independent, continuous speech decoding system has been developed to compare several training methods as well as several decoding algorithms within a whole integrated system. This system runs on a limited application, in French. Demisyllables are used as basic sub-word units. They are extracted from the application corpus semi-automatically. Speech units are modelled by different types of HMM-based techniques. A set of DTW/Viterbi-based decoding methods is available to perform recognition of sentences. The output of the decoding process can be either a lattice of syllabic units or a single string of decoded words. The lattice output is devoted to provide an AI speech understanding system with a suitable phonetic input. An evaluation module has been designed to evaluate the potentiality of the decoded lattices. The system is fully integrated in a software graphic interface which simplifies the management of the different modules and the understanding of the various outputs.