Estrategias de generación y reducción de variantes de pronunciación en sistemas de reconocimiento automático de habla: consideraciones arquitecturales

In the context of large vocabulary speech recognition systems, it is crucial to accurately model the allophonic variations that will be found in a real world task. In this paper we describe a study on the use of data driven pronunciation variations, considering the generation and reduction strategies, as well as their impact in the system performance. The described techniques are supported by the corresponding experimental evaluation on two radically different systems in what respect to their discrimination power (based on integrated and non-integrated architectures, designed to work as hypothesis and verification modules, respectively), so that it’s possible to discuss on their relative performance as a function of the increase in dictionary size. The most relevant results show that in the case of the non integrated architecture, we can significantly improve the inclusion rate, even for huge increases in dictionary size (up to 250%). On the contrary, the increase in the number of pronunciation variants has a clearly negative effect when applied to the integrated system.