The performance of a recognizer or recognition algorithm is studied as a function of the variation of specific speech production and speech transmission parameters (production and post-production factors). Production factors include all speaker related production parameters. Post-production factors include speech transmission parameters. The method uses a data-base with minimal difference word sets based on CVC-words. The CVC-words are embedded in a carrier phrase in order to obtain connected word conditions. Three groups of test-words are used, related to: initial consonants, final consonants and vowels. By means of an analysis-resynthesis technique the words are physically manipulated in such a way that the effect of the manipulations corresponds to a range of conditions for a given parameter representing variations within and between male and female voices. This method was presented earlier by Steeneken and v. Velden (1989). The extension to connected word recognition is described in this paper.
Bibliographic reference. Steeneken, Herman J. M. / Velden, Jeroen G. van (1991): "Ramos - recognizer assessment by means of manipulation of speech applied to connected speech recognition", In EUROSPEECH-1991, 529-532.