This study examined factors affecting perception training of vowel duration in L2 Japanese with transfer to production. In a pre-test, training, post-test design, 48 L1 English speakers were assigned to one of three groups: auditory-visual (AV) training using waveform displays, auditory-only (A-only), or no training. Within-group variables were vowel, preceding consonant, pitch pattern, and training talker's voice. Perception pre- and post-tests measured identification accuracy and response time (RT). Training involved eight sessions with feedback, including waveforms for the AV group. Results indicated significant improvement for the AV and A-only groups with generalization to novel stimuli and a new voice as well as transfer to production; the AV group showed a greater rate of improvement. Participants found waveform displays very helpful. Vowel type, preceding consonant, and pitch pattern significantly affected perception in testing and training as did the training talker's voice. The easiest pitch pattern was Low-High in the first syllable, perhaps reflecting English prosodic preference, and High--High in the second, which may be more salient. Perception was facilitated by talkers demonstrating greater pitch movement. Accuracy and RTs increased after training; participants reported spending more time evaluating post-training input. Results support the perception-production link, and the role of variable talker- and context-dependent perceptual categories.