VibeVoice-ASR-Streaming releases weights for speaker-attributed transcription
The streaming ASR model interleaves audio chunks, lookahead audio, and earlier text to transcribe who said what in near real time. The release includes 1.5B and 7B weights and inference code. The technical report says the 7B model has the lowest average WER/CER across five evaluation sets and the best or tied-best speaker attribution in 12 of 13 settings.
Visit original source