Session 4: Speech I

The first paper, "Field Test Evaluations and Optimizations of Speaker Independent Speech Recognition for Telephone Applications," by Gagnoulet and Sorin of CNET, was presented by Christel Sorin. This paper discussed various ways of improving system usability and performance by optimizing both the dialog ergonomy and the recognition technology within the constraints of low-cost real-time implementation. Techniques discussed included use of field data in training, increasing the number of parameters, automatic adjustments of the HMM structure, and better rejection procedures. A brief discussion of the rejection rate versus error rate tradeoff ensued; nobody had any good data or ideas on how to make this tradeoff, so when one person suggested that the rejection rate should be adjusted to keep the error rate under 5 percent, we said OK and moved on.