A new kind of feature for speech recognition, called Bi-Mel-scale Frequency Cepstrum (BMFC) is proposed. To calculate the BMFC, the speech signal is first segmented into short intervals and the bispectrum of each segm...A new kind of feature for speech recognition, called Bi-Mel-scale Frequency Cepstrum (BMFC) is proposed. To calculate the BMFC, the speech signal is first segmented into short intervals and the bispectrum of each segment is estimated; then the bispectrum is smoothed by two-dimensional mel-scale inverse filter bank; finally, the bi-mel-scale frequency cepstrum coefficients are obtained by decorrelating the outputs of the filter bank with twodimensional Discrete Cosine Transform (DCT). Preliminary experiments show that the new feature can improve the performance of a telephone speech recognizer and is more robust to white noise than the conventional LPCC (Linear Prediction Coefficients) and MFCC (Mel-scale Frequency Cepstrum Coefficents) used in speech recognition展开更多
文摘A new kind of feature for speech recognition, called Bi-Mel-scale Frequency Cepstrum (BMFC) is proposed. To calculate the BMFC, the speech signal is first segmented into short intervals and the bispectrum of each segment is estimated; then the bispectrum is smoothed by two-dimensional mel-scale inverse filter bank; finally, the bi-mel-scale frequency cepstrum coefficients are obtained by decorrelating the outputs of the filter bank with twodimensional Discrete Cosine Transform (DCT). Preliminary experiments show that the new feature can improve the performance of a telephone speech recognizer and is more robust to white noise than the conventional LPCC (Linear Prediction Coefficients) and MFCC (Mel-scale Frequency Cepstrum Coefficents) used in speech recognition
