On the Technology of Linear Prediction for Speech Recognition
Ran Chong-sen · Computer Engineering and Science · 2004
This paper proposes a linear predictive analysis method.Spectral envelopes are obtained from estimating frequency samples,and frequency normalization can be applied to estimating spectral envelopes. The envelopes are on the mel-frequency scale and IDFT is applied to extracting sample autocorrelations. We finally obtain cepstral coefficients (SEC) from sampling autocorrelation results. HMM experiments show that SEC improves the recognition performance at low SNR.