DuplexDrama Dataset for Full-Duplex Spoken Dialogue Training
September 14, 2026
DuplexDrama provides 2,000 hours of synthesized audio featuring 64 voices across 13 personas. The dataset includes full-duplex behaviors like interruptions and backchannels, with 3.8% of turns exhibiting these characteristics to improve conversational realism.
HOW THIS AFFECTS YOU
●
builderThis provides the data needed to move beyond turn-based voice bots toward natural speech interaction.
●
researcherYou can use this to train models on realistic turn-taking and interruption logic.