How does the emotional tone (happy or sad) and gender (male or female) of the instructor's computer-generated voice in a multimedia lesson affect learning processes and outcomes for female learners (Experiment 1) and male learners (Experiment 2)? In two experiments, college students viewed a 2.5 min narrated slideshow on the formation of lightning and then took a posttest. In a 2 (emotion) × 2 (gender) between-subjects design, the instructor's computer-generated voice was either happy female, sad female, happy male, or sad male, as produced by text-to-speech software. Female students (in Experiment 1) who received happy instructors gave higher ratings of instructor happiness, their own felt happiness, and how engaging and facilitative the instructor was, indicating that they could perceive, feel, and socially relate to the voice's emotion, and these effects were stronger for the female voice than the male voice. Male students (in Experiment 2) showed the same pattern for perceived emotion, partially for felt emotion, and not for social connection with the instructor. In terms of learning outcomes, happy instructors produced better transfer test scores only for males and did not produce better retention test scores for males or females. These results provide partial support for the "positivity principle," which posits that people learn better from happy instructors than from sad instructors and indicate that the instructor's gender plays a moderating role.