Prsentation PowerPoint

1
COMPANDED LATTICE VQ FOR EFFICIENT PARAMATRIC LPC
QUANTIZATION
Stéphane Ragot France Télécom RD Lannion, France
Roch Lefebvre, Marie Oger University of
Sherbrooke Sherbrooke, Quebec, Canada

• INTRODUCTION

2. GMM-BASED LSF QUANTIZATION
The p.d.f of LSF vectors x in dimension n can be
modeled by a Gaussian mixture model of order M
given by
A parametric approach based on Gaussian mixture
models (GMM) has been recently developed for the
vector quantization (VQ) of linear-predictive
coding (LPC) parameters. Although the coding
performance is limited by the accuracy of the
underlying p.d.f.source model, this approach has
some interesting features, such as asymptotic
bit-rate savings, bit-rate scalability and
complexity independent of bit rate. A new
technique of ellipsoidal lattice
vector quantization (VQ) is described, based on
1) scalar companding optimized for Gaussian
random variables and 2) rectangular lattice
codebooks with fast trellis-based nearest
neighbor search. The Barnes-Wall lattice in
dimension 16 is applied to quantize the line
spectrum frequencies (LSF) of wideband speech
signals. The LSF are computed in a manner similar
to the AMR-WB speech coding algorithm. The
performance of memoryless and predictive LSF
quantization for different GMM orders (4, 8 and
16) is evaluated at 36 and 46 bits per frame.
Where
and
Example of GMM for 2-D LSF
3. MEAN-REMOVED KLT CODING AND NON-UNIFORM SCALAR
QUANTIZATION
Mean-removed KLT coding
The candidate is the representative of x in
the i-th GMM component. The candidates are
computed by mean-removed Karhunen-Loeve Transform
(KLT) coding, which is known to be optimal for
the quantization of correlated Gaussian sources.
The matrix is the KLT matrix given by The
non-uniform scalar quantization presented here
has interesting features bit rate scalability
and a complexity independent of bit rate.
Non uniform scalar quantization
2
COMPANDED LATTICE VQ
Stéphane Ragot France Télécom RD Lannion, France
Roch Lefebvre, Marie Oger University of
Sherbrooke Sherbrooke, Quebec, Canada
4. A TECHNIQUE OF COMPANDED LATTICE VQ
Binary lattices and error-correcting codes
Example of 2-D Lattices
A lattice is a set of discrete points defined by
In our work we will restrict ourselves to binary
lattices, which can be decomposed as
Companded lattice VQ
In practice, we will use the so-called
Barnes-Wall lattice defined by
The technique described for the companded scalar
quantization can be extended as follows.
Rectangular lattice VQ is applied to the vector d
instead of scalar quantization. The number of
quantization levels is given by
5. EXPERIMENTAL RESULTS
Results for memoryless GMM-based LSF quantization
and histograms of SD
Experimental setup

• The database is the NTT-AT wideband speech
  material (sampled at 16 kHz) which is
  multilingual, multi-speaker and lasts 5 hours
• The training database is generate by
  607,386 LSF vectors.
• The test database has 202,112 LSF vectors.
• The E-M algorithm is applied to the training
  database to estimate the GMM parameters for an
  order M4, 8 and 16. Memoryless and AR(1)
  predictive GMM-based VQ are tested.

Results for AR(1) predictive LSF quantization and
histograms of SD
Spectral distortion statistics
The performance of LSF quantization is evaluated
with the spectral distortion (SD). Two bits rates
(36 and 46 bits) and different GMM orders (M4, 8
and 16) are tested. The histograms of SD are
provided for 46 bits and 16 Gaussians.
7. CONCLUSIONS
6. MEMORY/DISTORTION COMPROMISE
We presented a new technique of companded lattice
VQ. The results show that the spectral distortion
of GMM-based VQ can be reduced by using
rectangular lattice codebooks instead of scalar
codebooks. The storage cost of the GMM parameters
increases with the numbers of Gaussians. With 4
or 8 Gaussians the storage required is below that
of AMR-WB with 46 and 36 bits with equivalent or
better performance for spectral distortion .

When the GMM order is growing the storage is
increasing. The gain in Spectral Distortion is
not so important, so the best compromise seams to
be 8 Gaussians. The storage required in AMR-WB
for the 36 and 46 bit LSF codebooks is 27.5
Kbytes
Memory/Distortion