VibeVoice launches as an open-source TTS and ASR model family featuring 60-minute single-pass transcription with speaker diarization, timestamps, and hotword support.
VibeVoice released an open-source family of voice AI models covering both Text-to-Speech and Automatic Speech Recognition. The ASR model processes up to 60 minutes of continuous audio in a single pass within a 64K token context, outputting structured transcriptions with speaker identity, timestamps, and content. The system uses continuous speech tokenizers running at an ultra-low 7.5 Hz frame rate and a next-token diffusion framework combining an LLM backbone with a diffusion head. The release is currently positioned for research and development use only, with commercial deployment discouraged pending further testing.
The 7.5 Hz continuous tokenizer architecture is the real story here — it dramatically reduces sequence length for long audio without chunking, eliminating the speaker-tracking drift that plagues every chunked ASR pipeline. The joint ASR + diarization + timestamping in a single forward pass means you can replace three separate model calls with one. The diffusion head for TTS is also architecturally novel — this isn't a fine-tuned Whisper wrapper.
Pull the VibeVoice-ASR model from the project repo this week and run it against a 30-minute internal meeting recording — benchmark WER and diarization accuracy against your current Whisper + pyannote stack to get a real substitution signal.
Clone the VibeVoice repo from the project page and install dependencies with: pip install -r requirements.txt
Tags