First International Conference on Spoken Language Processing (ICSLP 90)
This paper presents an approximation to the recognition of Spanish digits based on a phonetic approach using probabilistic functions of Markov chains. The way to incorporate the phonetic knowledge to the markovian modelling resides on the strong sensitivity of the models to the initial conditions. Continuous PDF's, associated with each state allowed to represent the segments of least variability with greater weight than the others, and in the recognition stage assign more discriminative power to the phonetic relevant segments. The material under study consisted of the ten Spanish digits recorded from two utterances of 28 native male speakers, for a total of 560 words to recognize. The recognition accuracy of 99.64% confirm that the use of phonetic knowledge in the training step of the system allowed to define high discrimination states which favored the recognition. This type of training allowed to obtain an attractive relation of performance vs. complexity of the model.
Bibliographic reference. Gurlekian, Jorge A. / Franco, Horacio E. / Santagada, Miguel (1990): "Speaker independent recognition of isolated Spanish digits", In ICSLP-1990, 529-532.