Algorithms and Methods for the Automatic Speech Recognition in Spanish Language using Syllables

This work examines the results of incorporating into Automatic Speech Recognition the syllable units for the Spanish language. Because of the boundaries between phonemes-like units its often difficult to elicit them; the use of these has not reached a good performance in Automatic Speech Recognition. In the course of the developing the experiments three approaches for the segmentation task were examined: a) the using of the Short Term Total Energy Function, b) the Energy Function of the Cepstral High Frequency (named ERO parameter), and c) a Knowledge Based System. They represent the most important contributions of this work; they showed good results for the Continuous and Discontinuous speech corpus developed in laboratory. The Knowledge Based System and Short Term Total Energy Function were used in a digit corpus where the results achieved using Short Term Total Energy Function alone reached 90.58% recognition rate. When Short Term Total Energy Function and RO parameters were used a 94.70% recognition rate was achieved. Otherwise, in the continuous speech corpus created in the laboratory the results achieved a 78.5% recognition rate using Short Term Total Energy Function and Knowledge Based System, and 80.5% recognition rate using the three approaches mentioned above. The bigram model language and Continuous Density Hidden Markov Models with three and five states incorporating three Gaussian Mixtures for state were implemented. By further including a major number of digital filters and Artificial Intelligent techniques in the training and recognition stages respectively the results can be improved even more. This research showed the potential of the syllabic unit paradigm for the Automatic Speech Recognition for the Spanish language. Finally, the inference rules in the Knowledge Based System associated with rules for splitting words in syllables in the cited language were created.