This paper describes the use of Recogniser Sensitivity Analysis (RSA) for performance assessment of a connected word, speaker independent speech recogniser. A number of speech-production factors were considered and an analysis of variance indicated that relative energy, word duration and speaker consistency were found to account for over 86% of the variability in recognition performance for this data set. Second-order interaction effects between these were not significant indicating that the parameters can be considered independent. This permits the performance to be predicted as a function of each parameter independently. The results confirm the value of RSA both for obtaining diagnostic information about recognition performance, and as a method of predicting performance. Keywords: performance assessment; continuous speech recognition; recogniser sensitivity analysis
Bibliographic reference. Winski, Richard / Kordi, Kamran (1991): "Assessment of continuous speech recognisers using recogniser sensitivity analysis", In EUROSPEECH-1991, 521-524.