We report a complete evaluation of a Spanish Inferred Stochastic Regular Grammar (ISRG)-based Continuous Acoustic-Phonetic Decoder. We designed three basic series of experiments on a standard Spanish database which involves 1700 sentences uttered by ten speakers. For each one of these series we report results by using different feature analyses and sub-word language models. The different improvements have lead to the best performances of the system: 78% for the Speaker Dependent experiment and 72% for the Speaker Independent experiments. Moreover, the introduced benchmarks are not only a baseline evaluation for Spanish phone recognition but also allow for a direct comparison of ISRG with the more "standard" Multiple-Codebook Semi-Continuous Hidden Markov Models (MCSCHMM)for the same task and unit baseline.

Keywords: Syntactic Pattern Recognition, Grammatical Inference, Acoustic-Phonetic Decoding.