Evaluation of an isolated word recognizer in talker-dependent and talker-independent modes using a large telephone-band data base

An isolated word recognizer has been evaluated using a large data base of telephone-band digit utterances recorded by 100 talkers. Three reference template configurations of the recognizer have been studied, one talker independent and two talker dependent. The talker dependent configurations are a 1 template per word system obtained by robust training and a 5 template per word system. Performance has been studied in terms of the distribution of error rates across the three template configurations, the effect on performance of varying a decision rule parameter, applying a rejection threshold, and normalizing test and reference utterance lengths has also been investigated. Overall average error rates obtained are 2.17% for the talker independent system and 2.77% and 0.77% for the 1-template and 5-template talker dependent systems, respectively.