Spanish phone recognition using semicontinuous hidden Markov models

Abstract

The design of current acoustic-phonetic decoders for a specific language involves the selection of an adequate set of sub-lexical units and the choice of the mathematical framework in order to model such units.This paper deals with both items. On the one hand, we will discuss the choice of sub-lexical units we made for a Spanish continuous speech decoder. Our goal was to build a general, vocabulary-independent, speaker independent Spanish Continuous Speech phonetic decoder with only sublexical knowledge. For this purpose, we used a corpus of 1700 Spanish sentences uttered by 10 speakers.On the other hand and within the framework of the Semicontinuous Hidden Markov Modeling, we will discuss different approaches to the Viterbi-based re-estimation procedure. In this case we aimed to take advantage of the jointly optimization of the codebook together with the parameters of the model as well as reducing the computation cost of updating the codebook in the training phase.* This work has been partially supported by grants TIC 448/89 and TIC 1026/92-C02 of the Spanish CICYT. 11-5150-7803-0946-4193 $3.00 0 1993 IEEE

Integrating different acoustic and syntactic language models in a continuous speech recognition system

“…But, to increase the accuracy of the acoustic models in the CSR system, two different ways were considered: a) The use of a semicontinuous HMM for each phone-like unit instead of the discrete one. The joint optimization of the codebook together with the parameters of the model is one of the main advantages of this kind of models [5]. b) To increase the number of sub-lexical units keeping the use of discrete HMM.…”