Session 11: Continuous Speech Recognition And Evaluation I

This was the first of two companion sessions which marked an impor tant transit ion in the continuous speech recognition (CSR) component of the DARPA Spoken Language Program. Since 1987, DARPA CSR systems have been developed and evaluated on the Resource Management (RM) CSR corpus, which has become a de .facto standard for comparison of speech recognizers, widely accepted and used both within and outside the DARPA research community, and internationally. The RM corpus has served the community well, but due to its l imitations (e.g., 1000word vocabulary, restricted task domain), it has become clear that more challenging tasks were needed to drive CSR research. The papers in this session reported on the design and preliminary development and analysis of a new large-vocabulary CSR corpus for the 90s, which is aimed at meeting the needs of CSR research, and at serving in a complementary role to the corpora being collected in the interactive Air Travel Information System (ATIS) domain.