5th International Conference on Spoken Language Processing
This paper concerns the identification of Arabic macro-classes and phonetic features by systems using a hierarchy of neural networks. These systems are composed of sub-neural-networks (SNNs) carrying out binary discrimination sub-tasks. Two types of architecture are presented: serial structure of experts and parallel disposition of them. This mixture of experts is composed of typically time delay neural networks using a version of autoregressive backpropagation algorithm (AR-TDNN). These hierarchical configurations are confronted to a monolithic system using standard backpropagation learning procedure. The test database consists of 60 VCV utterances and 50 phrases pronounced by 6 Algerian native speakers. The parallel configuration achieved much fewer error rate (13% vs. 16% and 28%) than other architectures. The parallel mixture of experts is incorporated in a hybrid structure (HMM-SNN) in the order to enhance performances of standard HMMs. Identification results show that 10% reduction of error rate is obtained by the hybrid system.
Bibliographic reference. Selouani, Sid-Ahmed / Caelen, Jean (1998): "Modular connectionist systems for identifying complex arabic phonetic features", In ICSLP-1998, paper 0358.