Biomorphic Modeling Of Phoneme Identification And Classification Based On An Evolving Fuzzy-Neural Network From Hardcomputing To Softcomputing

Speech is dynamic in nature and organized in a complex time-and-frequency structure that makes it very hard to solve the issue of automatic speech recognition (ASR) for diverse speaker conditions. The hardcomputing approach to solving this issue (i.e conventional computing based on precisely-stated, analytical, mathematics-inspired models) pushed processing limits because it is highly computationally time-consuming and intolerant of imprecision, uncertainty or approximation in the data. Softcomputing and its biomorphic implementation is a more natural approach to solving the issue of speaker-independent ASR, given its ability to manage imprecision, uncertainty, and approximation, as well as to reduce system complexity to fit the upcoming requirements of next-generation deeply-embedded systems. This paper reports experiments based on an evolving-fuzzy-neural-network (EFuNN) paradigm trained to process and classify phonemes to drive multimodal (audiovisual) speech-to-text transcription and speaker identification