Matrix quantization error compensation for robust speech recognition based on neural networks

This paper proposes a robust, speaker-independent isolated word speech recognition (IWSR) system (MQ/HMM_VQ/HMM)/MLP which combines Dual Matrix Quantization (MQ) and Vector Quantization (VQ) pair combined with both the strength of HMM in modeling stochastic sequences and the non-linear classification capability of MLP neural networks. The system efficiently utilizes processing resources and improves speech recognition performance by using NN as the classifier of the system. Computer simulation clearly indicates superiority over conventional VQ/HMM and MQ/HMM systems with 98% and 93.8% recognition accuracy at 20 dB and 5 dB SNR levels, respectively in a car noise environment, based on the database TIDIGITS.