The BNTUW '98 workshop also differed in methodology from the previous benchmarks. The evaluation set was unsegmented, as compared to last year where both segmented and unsegmented speech was provided. This year a portion of the training and evaluation sets consisted of speeches recorded from CSPAN. Also, for the first time CSR systems were evaluated on Spanish and Mandarin broadcasts, and their performance was found to be comparable to English-language systems. The prevalent concern voiced at the workshop was the need for better metrics to evaluate speech recognition systems than the simple word error rate; as the focus shifts from speech-to-text conversion to speech understanding and information retrieval.

In this talk we present a brief review of the workshop and the systems which participated in the evaluations. We will also describe some of the most sophisticated information retrieval systems that were demonstrated during the course of the workshop. An exciting idea emerging out of these was the use of the Internet to perform remote speech collection and recognition, something of great interest to ISIP.