A hybrid HMM/autoregressive Time-Delay Neural Network Automatic Speech Recognition system

This paper describes a new hybrid approach which aims to significantly improve the performance of Automatic Speech Recognition (ASR) systems when they are confronted with complex phonetic features such as germination, stress or relevant lengthening of vowels. The underlying idea of this approach consists of dividing the global task of recognition into simple and well-defined sub-tasks and using hearing/perception-based cues. The sub-tasks are assigned to a set of suitable Time-Delay Neural Networks using an autoregressive version of the backpropagation algorithm (AR-TDNN). When they are incorporated in the hybrid structure, the AR-TDNN-based experts act as post-processors of a HMM-based system which thus acquires the ability to overcome failures due to complex language particularities. Results of experiments using either static or dynamic acoustic features show that the proposed HMM/AR-TDNN system outperforms that of the HMM-based system.