@inproceedings{, author={Kocsor, Andr{\'a}s and Kuba Andr{\'a}s Jr. and T{\'o}th L{\'a}szl{\'o}}, title={An Overview of the OASIS speech recognition project}, abstract={This paper presents an overview of the "OASIS" segment-based speech recognition project developed at the Research Group on Artificial Intelligence of the Hungarian Academy of Sciences. The aim of this project is to build a speech recognizer for Hungarian natural numbers. For this, a traditional spectral representation is computed first, from which acoustic-phonetic features are extracted. Some of these, like the energies of certain frequency bands were chosen in accordance with our knowledge about human auditory processing, while others, like "sonority" and "voicedness" measure the supposed acoustic correlates of these phonetic features. Based on the change of the features the speech signal is segmented into phonetically stable parts, and so-called interval-features are calculated over these segments. To each interval-feature belongs a function called "cue", which gets a segment and a phoneme as input and returns a punishment which indicates how probably the segment can be the given phoneme - according to that interval-feature. The cues are trained on the distribution of the feature on the database. The punishments of the several cues are aggregated by the phoneme evaluator, which thus can compute how well a phoneme fits to an interval (segment or series of segments). Finally, the matching engine matches all the possible segmentations to all the possible phoneme strings given by the dictionary, and returns the one with the smallest punishment as the result of recognition. In the past year we have examined many possible feature sets. We have also examined several methods for the phoneme evaluation, e.g. we used the C4.5 system and also tried an instance-based learning technique. The matching engine has also developed a lot: here pruning of the search space is crucial for an acceptable speed, but it is very difficult to find a proper aggregation and normalization of the punishments which allows the comparison of segment series of quite different lengths. Currently our system uses a set of 19 features, and gives the best results with the C4.5 evaluator and with the matching engine that traverses the search space with a backtrack algorithm. We give a detailed description of all these modules in the paper, and also present our recognition results on the phonemic and word level, trained on a database of 26-26 words from 20 talkers. }, booktitle={Proceedings of the 4th International Conference on Applied Informatics}, year={1999}, month={August}, address={Hungary}, editor={Eger-Noszvaj}, pages={94-102} }