Matrix quantization with vector quantization error compensation for robust speech recognition

Abstract

This paper proposes a robust, speaker-independent IWSR system which combines dual fuzzy matrix quantization (FMQ) and fuzzy vector quantization (FVQ) pairs, or dual MQ/VQ quantization pair with a discrete HMM to efficiently utilize processing resources and improve speech recognition performance. This system exploits the evolution of the speech short-term spectral envelopes with error compensation from FVQ/HMM, or VQ/HMM processes to target noise-affected input signal parameters and minimize noise influence. The enhanced processing technology employs a weighted LSP distance measure in the LBG algorithm. Computer simulation using gender-dependent HMMs clearly indicates the superiority over conventional FVQ/HMM and FMQ/HMM systems with 96.48% and 92.8% recognition accuracy at 20 dB and 5 dB SNR levels, respectively in a car noise environment, based on database TIDIGITS.