On the use of line spectral frequency parameters for speech recognition

The line spectral frequency (LSF) representation has been proposed by Itakura [l] as an alternative linear prediction (LP) parametric representation. In the context of speech coding, it has been shown [2-61 that this representation has better quantization properties than the other LP parametric representations (such as log area ratios and reflection coefficients). The LSF representation is capable of reducing the bit-rate by 25-30% for transmitting the LP information without degrading the quality of synthesized speech [4,5]. Our interest in LSF representation has been to see whether we can obtain a similar advantage from this representation for speech recognition. For this, we studied this representation in our earlier paper for the recognition of steady-state vowel frames in the speaker-dependent mode using the minimum distance classifier [7]. Though the LSF representation resulted in good performance [7], the scope of these results was very limited. The aim of the present paper is to extend the use of the LSF representation for more general speech recognition systems and to widen the scope of its results. (Some of these results have been reported earlier in a conference [8].) For this, we study here this representation in both the speaker-dependent and the speaker-independent modes for the hidden Markov model (HMM)-based isolated word recognition systems. Since the HMM-based speech recognizers use the maximum likelihood decision rule for recognition, we also report here the results for the speaker-&pendent and the speaker-independent vowel recognition experiments using the maximum likelihood classifier, In the present paper, we compare the performance of