Multi-Modal Signal Processing

Chapter | 3 Speech Processing

49

as context-oriented clustering (COC) [46]. COC builds a decision tree

for each phoneme, which automatically ‘explains’ the variability in

the acoustic realisation of this phoneme in terms of contextual factors.

Starting from a speech database with phonetic labels, it constitutes

initial clusters composed of all the speech units (allophones in this

case) with the same phonetic label (Figure 3.2). The context tree is

then automatically derived by a greedy algorithm. At each step, the

cluster with the highest variance (and with more than a minimum

number of segments) is split into two classes on the basis of a parti-

tion ...