BayLing-Duplex: Native Full-Duplex Speech Dialogue with a Single Autoregressive LLM

In three linesBayLing-Duplex is a native full-duplex speech language model using a single autoregressive LLM without external VAD module. Fine-tuned on 400K samples with DPO, it achieves 92% turn-taking success and 100% interruption success on InstructS2S-Eval, improving speech-response score from 2.17 to 3.39 over Moshi.

Summary generated by Claude — human-verified