Using linguistic knowledge to improve the design of low-bit rate LSF quantisation

In this paper we investigate an alternative approach to the design of low-bit rate (LBR) quantisation. This approach incorporates phonetic information into the structure of Line Spectral Frequency (LSF) codebooks. In prior work vector quantisation (VQ) has been used to quantise stochastic processes. Speech signals can, however, be described in terms of phonetic segments and linguistic rules. A trained LSF codebook, like the phonetic inventory of a language, is a static description of spectral behaviour of speech. As clear relationships exist between phonetic segments and LSFs the structure of an LSF codebook can be analysed in terms of the phonetic segments. The investigation leads to the conclusion that phonetic information can be usefully employed in codebook training in terms of perceptual performance and bit-rate reductions.