Development of a robust speaker-independent isolated-word recognition system using modular fuzzy neural networks.

摘要

In this investigation, a modular artificial neural network (ANN) architecture was designed that significantly reduced the training time of a conventional ANN based speaker-independent 21-word automatic speech recognition (ASR) system. The modular ANN architecture also improved recognition rates as compared to a conventional ASR system. The introduction of a gender neural network which distinguished a female voice from a male voice at the first stage of the modular architecture further enhanced the performance of the system. The resulting multistage modular artificial neural network (MMANN) based ASR system was built. The performance of this newly designed MMANN was tested with a speaker-independent recognition task of 21 isolated words, using a new set of speech data that was not used during the training process. The MMANN based speech recognition system demonstrated a performance improvement over systems based on the conventional Dynamic Time Warping (DTW) algorithm and a single monolithic ANN architecture. Fuzzy logic was combined with artificial neural networks to further improve the performance of the MMANN based word recognition system, and a multistage modular fuzzy neural network (MMFNN) based ASR system was developed. Fuzzy logic was applied to the 21-word recognition problem because it enables incomplete or imprecise data to be distinguished in a manner similar to that of humans. In the MMFNN architecture, proper use of fuzzifiers is critical to producing peak performance. The proper fuzzifiers help increase dissimilarity among the various classes in feature vectors. Empirical studies were conducted to obtain the proper fuzzy parameters since there was no general rule to decide the best values for the fuzzifiers. As a result, an MMFNN based ASR system was developed which performed best among several conventional and unconventional systems while reducing training time considerably. The modular architecture of the proposed system was also found suitable for concurrent training of individual neural networks. A very-large-scale integration (VLSI) implementation of the entire system should be useful for achieving real time performance in an isolated-word recognition system.