Comparison of parametric and non-parametric representations of speech for recognition

In this paper, we compare several feature sets based on the parametric and non-parametric representations of speech. Parametric representations are reflection coefficients, LPC derived cepstral coefficients (CCs) and line spectral frequencies (LSFs). Non-parametric representations are based on mel-frequency cepstral coefficients (MFCCs). These different representations are evaluated by their scores of recognition, for a speaker independent, isolated word recognizer based on hidden Markov models (HMMs).