Segment matrix vector quantization and fuzzy logic for isolated-word speech recognition

A novel speech recognition approach using segment matrix vector quantization (SMVQ) and fuzzy logic recognizer (FLR) is presented. SMVQ incorporates time sequence information and segment characteristics of speech signals. Firstly, the feature vector sequences of the speech signal is nonlinearly normalized to M frames. Secondly, the sequence is divided into N equal length segments. Finally, VQ is carried out separately and a codebook is designed for each segment. As a result, SMVQ can reduce quantization error and the sizes of codebooks. The subsequent recognizer using fuzzy logic technique with regards to each word to be recognized as a fuzzy set and conducts fuzzy reasoning. This recognizer need not process time alignment and complicated computations. It can be conveniently implemented in VLSI. For speaker independent isolated digit recognition, a recognition accuracy of 99.09% has been achieved by this approach.