A Distributed Architecture for Real-time Dialogue and On-task Learning of Efficient Co-operative Turn-taking - Coverbal Synchrony in Human-Machine Interaction - page 314

In the single-person condition, overlaps get continually fewer
(simple regression F = 3.39; p < 0.08), but improvement in reaction time
is not statistically significant although still indicative of the same trends
as observed in the other conditions. The observed improvements are
nevertheless in the expected direction, indicating that the system is, in
fact, improving during its interactions with this particular individual,
as it did with statistical significance for the more consistent artificial
interlocutor. In future we will seek to improve the performance for
individuals, since a noticeable adaptivity at the individual level is a
worthy, quite impressive goal to reach.

6.2 Silence tolerance window (STW) by pattern

We look for turn-giving intonation patterns in the last 300 msecs
of speech before each silence. Tail pattern of the pitch is currently
categorized into nine semantic categories based on slope (Up, At,
Down) and final pitch compared to average (Above, At, Below).
To begin with, we will analyze the distribution of these patterns
before the silences that mark end of turn, and before the silences
that are within turn. In both the artificial interviewee evaluation
and single-person evaluation, the pattern Down_Below (representing
a final fall in pitch) is most widely used at end of a turn (see Table
4). This harmonizes well with previous research (Pierrehumbert and
Hirschberg, 1990), which has associated a final fall in pitch with a turn-
giving signal. Furthermore, the person and the artificial interviewee
have a very similar distribution of patterns at the end of a turn. The
same cannot be said about prosody patterns perceived before silences
that do not lead to turn transition. Prosody before silences within

Table 4. Distribution between prosody patterns.

Pattern Artifi cial Interlocutor Human Interlocutor

At end Within At end Within

Down_Below

58.6% 0.4% 42.0% 12.6%

Straight_Below

10.3% 0.1% 14.1% 17.2%

Up_Below 8.6% 2.4% 7.3% 3.6%

Down_At 8.4% 20.6% 10.4% 14.6%

Up_Above 5.6% 38.1% 5.0% 14.4%

Straight_At 2.7% 10.6% 6.5% 15.7%

Straight_Above 2.2% 13.7% 7.3% 9.7%

Down_Above 2.1% 10.2% 3.4% 4.6%

Up_At

1.5% 4.1% 3.9% 7.6%