From hybrid adjustable neuro-fuzzy systems to adaptive connectionist-based systems for phoneme and word recognition

Abstract

This paper discusses the problem of adaptation in automatic speech recognition systems (ASRS) and suggests several strategies for adaptation in a modular architecture for speech recognition. The architecture allows for adaptation at different levels of the recognition process, where modules can be adapted individually based on their performance and the performance of the whole system. Two realisations of this architecture are presented along with experimental results from small-scale experiments. The first realisation is a hybrid system for speaker-independent phoneme-based spoken word recognition, consisting of neural networks for recognising English phonemes and fuzzy systems for modelling acoustic and linguistic knowledge. This system is adjustable by additional training of individual neural network modules and tuning the fuzzy systems. The increased accuracy of the recognition through appropriate adjustment is also discussed. The second realisation of the architecture is a connectionist system that uses fuzzy neural networks FuNNs to accommodate both a prior linguistic knowledge and data from a speech corpus. A method for on-line adaptation of FuNNs is also presented.