Efficient LPC Quantization Methods - Digital Speech: Coding for Low Bit Rate Communication Systems - page 132

1000

f 1
g 1

800

600

400

200

0

0 500 1000 1500 2000

Time (samples)

Figure 5.20 LSF tracks f 1 and g 1 variations over time

method due to the weak stationarity assumption within the analysis window,
especially at transitions from voiced speech segments to unvoiced (offsets)
and vice versa (onsets). The low-pass filtered method, on the other hand,
produces smoother and more slowly evolving LSF tracks. The differences in
the LSF tracks are more evident in the higher LSF parameters ( f 7 and f 10 )as
shown in Figures 5.22 and 5.23.
Work in [16] showed that using a perceptually-smoothed power spectral
envelope leads to a significant increase in subjective performance. Addition-
ally, [17] showed that low-rate quantization is possible through smoothing
the LSF parameter evolution. An informal listening test comparing both the
classic, f , and low-pass filtered, g ,LSFvectorsusedina4kb/sSB-LPCcoder
showed no difference in speech quality. An advantage during quantization is
therefore expected with regard to smoother evolution of the LSF tracks.

5.10.2 AdvantagesofLow-passFilteringinMovingAveragePrediction
Although using the unquantized LSF parameters for both the new and classi-
cal methods did not show any subjective quality difference, the new method
is expected to produce better performance under predictive quantization.