Abstract Children who use cochlear implants (CI) and who are readers usually produce more accurate speech in response to text than to pictures. Equivalence-based instruction (EBI) can be a route to establish functional interdependence between these verbal operants. The present study investigated whether children with CI who read would improve speech accuracy when tacting pictures of scenes after EBI that included dictated sentences, pictures of scenes, and printed sentences. This study evaluated whether teaching verbal relations to diagonal sentences from a matrix with subject-verb-object combinations promoted recombinative generalization to untrained sentences. Participants were three children with CI with a more accurate speech when reading print than when tacting pictures of scenes. They were taught to select pictures of scenes in response to dictated sentences (AB) by matching-to-sample (MTS) and to construct printed sentences in response to dictated sentences (AE) by constructed-response-matching-to-sample (CRMTS). Speech production in response to print (CD) and in response to pictures of scenes (BD) were probed for both trained and untrained sentences, using a multiple baseline design across participants. All participants learned the trained relations, showed emergence of derived relations, and improved speech accuracy when tacting pictures of scenes. They were able to recombine sentence components and tact novel pictures using untrained sentences from the matrix. These results indicate that speech accuracy and generative sentence production can be improved in children with CI from interventions that incorporate EBI and matrix training. Trial registration CAAE#01454412.0.0000.5441 registered 01/29/2013.