We investigated the effects of generating words from fragments on pronunciation time, on immediate memory span, and on delayed free recall. Subjects read long words and short words aloud or generated them from strings with missing letters. Word-length and generation condition had multiplicative effects on speaking rate, as expected if each affected a separate process regulating the rate. We replicated the standard finding that span is smaller for longer words. Generation improved delayed free recall, indicating that relatively brief presentation times are adequate to produce a generation effect. Although generation improved long-term memory for the words, memory span was shorter for the words that were generated. The harmful effect of generation on span appears to be due to its slowing of speaking rate.