next up previous
Next: Part-of-Speech Tagging Up: Data and Evaluation Previous: Performance Criteria

Testing Methodology

While some algorithms in this field are portrayed as multi-purpose systems which could be used in speech synthesis, speech recognition or automatic database tagging [Ostendorf and Veilleux, 1994], [Wang and Hirschberg, 1992], our system is designed and optimised solely for use in a unconstrained text-to-speech system. To that end we test the system on the input format it can expect in real use: a continuous stream of ascii characters with no division into words, sentences or other units. The division of the text into such units is performed automatically within the initial text processing modules of the text-to-speech system. The POS tagging is performed automatically, and no acoustic features are used. These testing conditions are as close as one can get to how the system would operate in real use: it is important to bear this in mind when comparing these results to other systems.

Alan W Black