VibeVoice-ASR-Streaming Technical Report
发表于Published in arXiv preprint, 2026

Traditional speaker-attributed ASR systems treat recognition and speaker diarization as two separate tasks. End-to-end models such as VibeVoice-ASR have since unified the two within a single model, but existing unified models mainly support offline recognition, which is hard to reconcile with the low-latency requirements of real-time voice assistants and agents. This report presents VibeVoice-ASR-Streaming, one of the first LLM-based end-to-end approaches to streaming speaker-attributed ASR. It interleaves fixed-size audio chunks, a small amount of lookahead audio, and previous text, allowing the model to produce “who said what” as speech arrives, without a separate diarization stage.
For transcription accuracy, the 7B model achieves the lowest average WER/CER across five evaluation sets. For speaker attribution, it is best or tied-best on 12 of 13 evaluation settings. The 1.5B and 7B weights are released together with inference code, including a vLLM plugin for serving.
推荐引用:Recommended citation: Y. Tu, Z. Peng, J. Yu, L. Dong, S. Xu, Y. Chang, W. Wang, Z. Wang, Z. Wang, Y. Xia, J. Zhang, X. Chen, and F. Wei. (2026). "VibeVoice-ASR-Streaming Technical Report." arXiv preprint.
