Both investment capital and development time can be reduced substantially if reusable tools and resources become available. Speech corpora to develop and train recognition systems are the most expensive and time consuming resource needed to build applications that use speech recognition in the PSTN. Moreover, annotated speech corpora are also indispensable for running objective performance tests of off-the-shelf technology that has proven adequate performance in some languages, but not yet in the language of the country of interest. Because large commercial interests are at stake with the introduction of automated services using ASR, one can understand that marketing managers are reluctant to accept new technologies as long as there is no definite proof of their adequacy for the home market.
In the fringe of ICSLP-'92, the international Cooperating Committee on Speech
Databases and Assessment (COCOSDA) defined guidelines
for a corpus that should be able to solve at least some of
the problems slowing down the uptake of speech technology applications.
Comparable corpora, now known under the name POLYPHONE, should be recorded
for as many different languages as
possible. In the course of 1993 these guidelines were made more concrete by
the Linguistic Data Consortium (LDC), which specified a recording protocol for
American English and Spanish [1,2].
Taking those protocols as a point of
departure, PTT Research and the Speech Processing Expertise
Centre SPEX set out to record the Dutch POLYPHONE corpus.