Modular connectionist systems for identifying complex arabic phonetic features

This paper presents an approach using a mixture of connectionist experts for the identification of complex Arabic phonetic features such as the emphasis, the gemination and the relevant duration of vowels. These experts are typically time delay neural networks using a version of autoregressive backpropagation algorithm (AR-TDNN). A serial and parallel architectures of AR-TDNN have been implemented and confronted to a monolithic system. The parallel configuration achieved much fewer error rate (13% vs. 16% and 28%) than other architectures. This leads us to develop a hybrid system based on hidden Markov models (HMM) and a parallel configuration of AR-TDNN. Binary discrimination sub-tasks are assigned to these neural experts in order to enhance HMM classification capabilities of complex phonemes. Results show that 10% reduction of error rate is obtained by the hybrid system in comparison with a baseline system.