SmarTspelLTM: a multipass recognition system for name retrieval over the telephone

We present a multipass recognition approach for continuously spelled name recognition. After describing this approach and presenting an optimization of the first pass using spectral parameter filtering techniques, we report on the evaluation of the whole system, both on a data base and during a four month field evaluation on two different tasks. The optimization of the first pass showed that relative spectral perceptual linear predictive (RASTA-PLP) parameterization is a good compromise between performance and speed. Evaluation on a data base yielded more than 95% name retrieval for a dictionary of 3388 names, while a field evaluation on small dictionaries (50 and 78 names) showed that nearly 90% of the calls were completed successfully. At the end of the paper, we discuss some of the insights provided by the field evaluation and show how this system can be extended to tasks with large dictionaries (on the order of 100000 names)