Reproduce “D2: High-Capability Full-Duplex Spoken Language Models from Half-Duplex Backbones” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cheng-Kuang Chang, Alexander H. Liu, Brian Kingsbury, Hilde Kuehne, Rogerio Feris, Samuel Thomas, James Glass, Kai-Wei Chang
D2 adapts pretrained half-duplex spoken-language models for full-duplex dialogue by distilling an offline audio encoder into a streaming encoder, then fine-tuning interaction in aligned audio chunks. At 80-ms chunks, Qwen3-Omni-D2 scores 70.1 on VoiceBench, 9.1 points above the evaluated full-duplex baseline MiniCPM-o 4.5. Its original backbone scores 86.8, so capability parity is not established.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.