This paper presents an approach to integrate frequential and temporal structurations of the speech signal in a symbolic learning and rule-based recognition process. This approach is evaluated on experiments of vowels recognition in continuous speech and compared with a statistical approach. To take into account the frequential and temporal structurations, we choose the Clustering of Spectral Peaks (CSP) measure based on peak parameters space and the Generalized Temporal Decomposition (GTD) modeling the spectral evolution. Charade is the chosen symbolic learning system. From our experiments, a priori strategic temporal information, such as localization or segmentation obtained from GTD technique using CSP measure, is not shown to be useful information for vowels identification. Our comparative study symbolic versus statistical approach is very encouraging for further research on symbolic process. Indeed, the results obtained from symbolic approach are quite comparable to the best results obtained from statistical approach.