Evaluation of the Telefonica I+D Natural Numbers Recognizer over different dialects of Spanish from Spain and America

Abstract

In this paper we present the results obtained when evaluating the Natural Numbers Recognizer of Telefónica I+D over some particular dialects of Spanish from Spain and America. The evaluation was made over two different data sets corresponding to two different situations. A first set includes dialects of Spanish from Spain, that were considered in the training and design of our baseline system, and a second set corresponds to Argentinian Spanish, that was not considered to train the original system. Just because we are interested in a system able to be used by a wide range of users, we tested the possibilities of MAP (Maximum-A-Priori techniques) to adapt the original HMMs in order to represent all the dialects. The experimental results show the capabilities of our recognizer to be used in applications spread over a great number of Spanishspeaking countries.

Multidialectal Spanish acoustic modeling for speech recognition

“…Research shows that a mismatch in dialects between training and testing speakers significantly influences recognition accuracy in several languages like French (Brousseau and Fox, 1992), Japanese (Kudo et al, 1996), Dutch (Diakolukas et al, 1997), German (Fischer et al, 1998) or English (Chengalvarayan, 2001), as an example. Spanish is not an exception, as it has been shown in research (de la Torre et al, 1996;Zissmanm et al, 1996;Aalburg and Hoege, 2003). Efforts in dialect ASR technology have followed two different goals: (i) to improve dialectal recognition rates by developing recognition systems tailored to specific dialects and (ii) to design multidialectal ASR systems robust to dialect variation.…”

“…Variability due to speakers and data from different dialects is considered to be pronunciation variation; as such, it is modeled by adding alternative pronunciations to the lexicon (Billa et al, 1997;Ferreiros and Pardo, 1999), or by defining a simple phonetic set (Huerta et al, 1997) in order to integrate variability in HMM. Two examples of specific dialectal modeling can be found in (Aalburg and Hoege, 2003;de la Torre et al, 1996). In the first paper, Spanish as spo-ken in Spain is used to model non-native speech applied to a system trained with Colombian speakers.…”

Reconocedor de habla continua independiente del contexto para el español de Argentina

“…Sin embargo, no se encuentran muchos trabajos sobre reconocimiento de habla continua en Español (Villarubia et al, 1997;Zhan et al, 1996) y en Español de Argentina (de la Torre et al, 1996) de los que se puedan comparar resultados sobre tasas de reconocimiento y tiempos de procesamiento, como los realizados para el idioma inglés (Ravinshakar, 1996) los cuales además de comparar dichas figuras de mérito adicionan la capacidad de memoria de almacenamiento de los modelos acústicos.…”