Sampling and Quantization - Digital Speech: Coding for Low Bit Rate Communication Systems - page 47

between the extreme cases of low computation cost with high storage (binary
codebook) and high computation cost with low storage requirement (full
search codebook).
During the training of a binary codebook, at each stage of splitting using
the K-means algorithm and method 1, the resultant optimum codebooks are
stored. The database is also split into sections represented by each of the
resultant vectors. When the vectors are further split, each new pair of vectors
is optimized using the section of the database represented by their mother
vector. This process continues until the final size codebook is reached and
optimized.

Cascaded Codebooks
The major advantage of a binary search codebook is the substantial decrease
in its computational cost, relative to a full search codebook, with a relatively
small decrease in performance. However, the storage required for a binary
search codebook relative to a full search codebook is nearlydoubled. Cascaded
vector quantization is a method intended to reduce storage as well as
computational costs [18, 13]. A two-stage cascaded vector quantization is
shown in Figure 3.12. Cascaded vector quantization consists of a sequence of
vector quantization stages, each operating on the error signal of the previous
stage. The input vector x is first quantized using a B 1 bit L 1 level vector

Vector Quantizer

Codebook 1 Codebook 2

y i

−

x(n) e(n) e

x

Vector
Matching

Vector
Buffer

Vector
Matching

Vector
Buffer

+

+

index i index k

Vector De-quantizer

index i

Codebook 1

^

x(n)

+

index k

Codebook 2

Figure 3.12 A two-stage cascaded vector quantizer