Three different geometric pattern recognition techniques for acoustic modeling of phone units (or other sub-word units) for continuous speech recognition are presented for their application to acoustic-phonetic decoding tasks. In the first methodology, each phone unit is modeled as a collection of templates in order to capture the variability of the acoustic events that characterize it, and the acoustic-phonetic decoding is performed by an improved version of the classical Two-Level algorithm. The other two methodologies are used integrated with hidden Markov models. The learning algorithms of the geometric components of such hybrid models allow to improve the discriminative skills of hidden Markov models, while the temporal deformation of the patterns is dealt with the Markov chain. Different speaker-independent and task-independent experiments with one Spanish database are performed.