Figure 4: Control and data flow between vocal and other behaviors in Kismet’s architec-
ture. Dark arrows indicate combined control and data flow between old components of the
architecture. Light arrows follow the data flows only.
5. Structures and Algorithms for Pro-
tolanguage Development
The development of the robot’s formulaic proto-
lexicon hinges on two mechanisms: one for the up-
date of the vocal label of an existing concept (or
“word”), the other one for the creation and man-
agement of novel concepts.
5.1 Acquisition of labels
The Concept class implements the functionality of a
single vocal behavior. Its overall architecture was
briefly explained in Figure 3, and its state machine
is presented in more detail in figure 5. After ini-
tialization, the behavior can be in one of the seven
states. On the figure, two non-default transitions
are shown from the Decide state: to Activate1 and
Hear states. They are determined as follows: after
the OutputLabel state has executed, the transition
goes to Activate1 unless the behavior receives a sig-
nal that there is a new speech input, in which case
the transition goes to Hear.
Figure 5: State machine of a Concept behav-
ior.
Activate1 computes the behavior activation. If
the activation is above a threshold, it transitions to
OutputLabel, whence the vocal label of the behav-
ior is sent out. Otherwise, the next state remains
the same, unless the condition for label update is
satisfied. Then the transition function leads to the
UpdateLabel state.
5.1.1 Attribution of label
When new speech input arrives, this activates the
Hear state of every Concept, which attempts to match
the heard phonemic string to the behavior’s vocal la-
bel. The match value is computed based on the vocal
label as a template against which to match, the con-
fidence value for the behavior’s own vocal label, and
an empirically determined global phoneme confusion
matrix:
BestM atch(in, template) = max(M (in, template)) (1)
where M (in, template) is a vector whose elements are
best matches given a certain window size (window
size is measured in phonemes and not in single char-
acters):
M (in,template) = —
n
Mm (in, template) • m
Mm +ι (in, template) • (m + 1)
• ∙ ∙
Mn (in, template) • n
(2)
where m is the minimum allowed window size and
n = min (MAX WINDOW SIZE,length (template))
The minimum window is set so that spurious one-
phoneme matches are discarded. The matching mea-
sures returned by Mi () are scaled by the window size
in order to bias in favor of longer substrings. Mi ()
itself uses the standard brute-force string searching
algorithm.
Hear automatically transitions to Activate2 which
computes the activation level of the behavior based
on the match determined earlier and the values of the
behavior’s receptors. If a discrepancy between the
response to receptor values and the response to the
speech input is above a threshold, and the confidence