A multiple-baseline technique was used to evaluate generalization effects during articulation training with trainable mentally retarded subjects. Four target words were selected for each subject on the basis of whether the subject could articulate the word correctly when it was modelled but could not articulate the word correctly in response to a picture of it. Five different settings were selected for generalization probing and training for each subject. In Setting 1, Experimenter 1 initiated training sequentially on all four target words for each subject. Other experimenters probed for correct articulation generalization in four other settings. Training was initiated in these four other settings sequentially only if correct responding failed to generalize to a setting. Results indicated that it was necessary to initiate training on at least three of the four selected target words in at least one additional setting with an additional trainer before correct responding generalized to untrained settings.