Milan Jelinek, Redwan Salami, Sassan Ahmadi, Bruno Bessette, Philippe Gournay
Milan Jelinek, Redwan Salami, Sassan Ahmadi, Bruno Bessette, Philippe Gournay
Milan Jelinek, Redwan Salami, Sassan Ahmadi, Bruno Bessette, Philippe Gournay
, Redwan Salami
, Sassan Ahmadi
*
, Bruno Bessette
, Philippe Gournay
, Claude Laflamme
\
|
=
o o
where o ( 1 0 s s o ) is a coefficient that controls the inter-
harmonic attenuation, T is the pitch period, x(n) is the lower
band of the reconstructed signal and y(n) is the post-processed
lower band signal. The factor o is derived from the normalized
correlation between the full band signal x(n) and its delayed
version 0.5{(x(n-T) + x(n+T)}. The band splitting filter is
combined with the upsampling filter to simplify the
implementation.
In the case of a WB output, the high frequencies are then
regenerated for active speech frames in the same way as in
AMR-WB and added to the upsampled post-processed synthesis
signal.
4.1 Frame erasure concealment
The erased frames are processed differently depending on the
last correctly received frame. If this frame happens to be a CNG
frame, no special processing is needed as the CNG parameters
are already interpolated. The only major difference is that the
decoding of these parameters is skipped and previous parameters
are used instead.
Processing of erased frames following an active speech frame is
done separately. The processing is however the same
independent of the coding type used in the preceding active
speech frame. The concealment is done by estimating the
excitation signal for the whole frame and filtering it with an
estimated LP synthesis filter, interpolated 4 times per frame as in
normal processing. The FER concealment can be summarized
such that the energy of the excitation signal and the spectral
envelope represented by the LP filter coefficients are gradually
moved to the corresponding estimated parameters of the
background noise. The excitation periodicity converges to zero.
The rate of the convergence depends on the last good frame
classification.
The second stage of the FER processing involves the
recovery of the normal processing after an erasure interval is
over. To improve the convergence to the normal operation when
a voiced onset has been lost, the onset is reconstructed
artificially to enable rapid synthesis convergence of the voiced
speech in the first good frame after the erasure and before
continuing with the ACELP