Improved line spectral frequency estimation through anti-aliasing filtering
A study of the classic line spectral frequency (LSF) extraction methods along with the assumptions made during their estimation is presented. LSF extraction is investigated from an over-sampling and decimation perspective and LSFs are shown to contain high-frequency variations that led to spectral overlapping problems. An anti-aliasing filter prior to decimation, with a cut-off frequency dependent on the final LSF vector transmission rate, is proposed to alleviate the aliasing problem of the classic extraction methods. The proposed method shows a clear advantage because it produces the same quantisation distortion as the classic methods at a lower bit requirement, with significant reduction in 2 and 4 dB outliers that greatly affect synthesised speech quality. This was also confirmed through a listening test.