Learning Strategies for Modular Neuro-Fuzzy Systems: A Case Study on Phoneme-Based Speech Recognition

Abstract

The paper discusses a general engineering model of a hybrid neuro-fuzzy system and its application to phoneme-based speech recognition and information retrieval. The speech recognition part consists of a low-level neural network module for phoneme recognition and a higher-level fuzzy reasoning module for word recognition and language modelling. Both low and high level modules are multi-modular, having a separate unit for each of the English phonemes. This architecture makes possible exploring different learning strategies to improve the recognition rate. The following ones have been introduced and experimented in the paper on the task of phoneme-based spoken digits recognition: i additional training/adjustment of individual neural network units; ii using both real and synthetic data in one training data set; iii averaging, over 3 time frames, the recognised speech signals; iv individual adjustment of membership functions in the fuzzy linguistic modelling units; v adjustment of a “level of tolerance” coefficient. It is demonstrated on a small scale experiment that all these strategies contribute to a better recognition rate of the overall system.