Sampling and Quantization - Digital Speech: Coding for Low Bit Rate Communication Systems - page 52

Vector Quantizer

Fixed
Codebook

Index k

x(n) x e

Vector
Matching

Vector
Buffer

+

y k

+

−

+

x p

Vector
Matching

Adaptive
Codebook

Index i

Vector De-quantizer

Index k

Fixed
Codebook

y k

^

x

x(n)

+

x p

Adaptive
Codebook

Index i

Figure 3.14 Adaptive vector quantizer in a cascaded setup

After designing a codebook to match a given set of training data, it is
important to test the performance of that codebook on data that was not
used in the training. Testing only on the training data will always give better
performance than the codebook will actually give in practice.
The robustness of a codebook can be measured by measuring its perfor-
mance on data whose distribution is different from that of the training data.
In practice, one cannot usually predict all of the situations under which a
quantizer will be used and so the distribution of the actual data may be
different from that of the training data. There are two major types of variation
that affect the design and operational performance of a codebook: input signal
variability and digital transmission channel errors.
Signal variability can be further classified as speaker variability and envi-
ronmental variability. Speaker variability covers the changes in the input
signal due to a change in the speaker's voice and may, for example, be due
to multiple speakers or the health conditions of each speaker. Environmental
variability, on the other hand, refers to the background noise level and type.
For a given bit rate and speaker, a speaker-dependent codebook performs