Statistical features versus word templates for speaker independent digit recognition over long distance telephone connections

First we extend to speaker independent recognition of isolated digits over long distance telephone connections (from the FAA telephone speech data base) the use of frame-specific statistical features that have already been tested [1,2] on a studio-quality data base (Texas Instruments isolated digit data base). In addition, for both the telephone and the studio-quality data bases, we compare recognition performance of frame-specific statistical models versus pattern matching of the input speech with multiple templates of the vocabulary digits. When testing is performed on the more difficult long distance telephone data base, the performance of frame-specific distance measures (3.6% substitution rate ) is superior to pattern matching with multiple templates (6.8% substitution rate).