Decision trees for inter-word context dependencies in Spanish continuous speech recognition tasks

Abstract

Context Dependent Units are broadly used in Continuous Speech Recognition (CSR) system, being decision trees a suitable clustering technique to obtain this kind of units. This work was aimed to extend the decision tree based clustering to model inter-word context dependencies in Spanish CSR tasks. We first used a set of previously defined context dependent units to model word boundaries. A decision tree derived pair grammar was then used at decoding time to prune each network connecting pairs of words. Then, specific sets of decision tree based inner context dependent units were obtained to model word boundaries. Both approaches were experimentally evaluated and compared to classical approaches over a Spanish CSR task. Experimental results showed the potential contribution of modelling between-word contexts to CSR systems. These units were selected by decision trees and provided full coverage while keeping a suitable computational cost.