A connectionist model for speaker-independent isolated word recognition
Mingming Zhu, Klaus Fellbaum · International Conference on Acoustics, Speech, and Signal Processing · 2002
A connectionist model for speaker-independent isolated word recognition is presented. It consists of an efficient preprocessor and a network of multilayer perceptrons. The preprocessor measures the local spectral similarities by implementing a multicodebook vector quantizer and compresses nonlinearly each speech pattern into a fixed length of 24 triples. The network makes the final decision by decoding these triples which contain temporal information and minimum distortions. Preliminary experiments show a recognition rate of 95.5%, which indicates that a properly designed combination of a preprocessing scheme and a neural network can greatly reduce the computational load as well as increase recognition rates.>