Abstract
This thesis analyzes and improves on the accuracy of the MFCCs (Mel-Frequency Cepstral Coefficients) feature extraction currently used in our lab’s fixed-point ASR (automatic speech recognition) system based on HTK (hidden Markov model toolkit). We propose three methods for improvement: first by changing the FFT (fast Fourier transform) algorithm, then by using a logarithmic representation for the power spectrum after FFT and the Mel-filter bank, and lastly we improve the method for multiplication by using double-length integers to achieve higher precision.Experimental results shows that each of the above methods yields an improvement in both fixed-point computation precision and recognition rates (by 2~3%) over the original fixed-point system. Further experiments on the effects of overflow at the Viterbi stage and background noise show no correlation of these effects with recognition rates.