Progress in Speech Recognition for Romanian Language
Corneliu-Octavian Dumitru, Inge Gavăt · InTech eBooks · 2008
In this chapter we presented the work done until now to implement a tool for research in speech recognition for Romanian language. What we realized can be summarized as follows: ? We have implemented basic algorithms for feature extraction for perceptual and simple cepstral analysis and perceptual and simple linear prediction ? We have implemented learning strategies based on HMMs, ANNs, and hybrid strategies like fuzzy-MLP, fuzzy HMM or HMM-MLP in the framework of the statistical, connectionist and fuzzy paradigms of Computational Intelligence ? We have constructed databases to train the models and to test our recognizers that obey the same statistical rules as given from linguists for Romanian language. ? We have trained the models and tested the recognizers on this data bases What we achieved and what is further to be done is sketched bellow: ? The obtained results in ASR for WRR are acceptable, what can not be said for ASRU experiments, so that much work must be done to enhance PRR by extending the knowledge resources to the semantic and pragmatic levels. Our resources for the moment cover the acoustic and phonetic level by the trained models and the dictionary and in a very primitive manner the syntactic level, by a loop grammar. A more complex grammar and a language model are our next objectives