VibeVoice-ASR-Streaming Technical Report
The paper puts speaker-attributed transcription into a streaming LLM pipeline.
VibeVoice-ASR-Streaming interleaves fixed-size audio chunks, limited lookahead, and prior text so it can output “who said what” as speech arrives. The authors frame it as one of the first LLM-based end-to-end systems for streaming speaker-attributed ASR, avoiding a separate diarization stage. They report that the 7B model has the lowest average WER/CER across five evaluation sets and best or tied-best speaker attribution in 12 of 13 settings. The team is releasing 1.5B and 7B weights with inference code. HF Daily Papers' note
VibeVoice-ASR-Streaming interleaves fixed-size audio chunks, limited lookahead, and prior text so it can output “who said what” as speech arrives. The authors frame it as one of the first LLM-based end-to-end systems for streaming speaker-attributed ASR, avoiding a separate diarization stage. They report that the 7B model has the lowest average WER/CER across five evaluation sets and best or tied-best speaker attribution in 12 of 13 settings. The team is releasing 1.5B and 7B weights with inference code. HF Daily Papers' note
score 5