- A software tool for analysis of
interaction between recognition performance and
test data
characteristics. Examples of data characteristics are speaker
characteristics such as sex, age, dialectal categories,
educational level, weight, etc.
- Other types of characteristics
could be recogniser type, testing-site, microphone,
testing
environment, etc.
- All these examples are that they can easily be represented
by subsets/sessions within the SAM_SCOR software, so that test
performance can be automatically derived for each subset/session
by means of SAM_SCOR.
-
Subsets/sessions are to be defined in the SAM_SCOR control
file
which can easily be set up using the SAM tool RISE.
- Input to SAMITPRO is provided by applying the SAM_SCOR to
a
series of result files from recogniser experiment(s), so that
the
performance is sampled in each cell of the input dimension.
- SAM_SCOR outputs the performance for each subset/session
in a
DBMS file, which can subsequently be read by SAMITPRO.
- SAMITPRO places the recognition performance in cells in
a
multi-dimensional table. Each dimension represents one characteristic,
i.e. cells are described by a number of categories that are
mutually exclusive. After filling up the table, hypotheses
concerning the interactions of data characteristics
(dimensions)
can be evaluated by means of log-linear model fitting.
From this analysis conclusions can be drawn concerning the
correlation between speaker characteristics and recognition
performance: for instance whether some characteristics
have an
impact on performance or whether a complex interaction exists
between several characteristics in combination.
Further work aims at extending SAMITPRO to provide the capability
of correlating recognition performance with speaker
dependent
speech inherent parameters derived by SAM_SPEX.
