One repository, three different availability stories
VibeVoice presents itself as a family of voice models rather than one package with one job. The most complete current path is VibeVoice-ASR, a 7B speech-recognition model designed to accept up to 60 minutes of audio in one pass. It produces a structured transcript that combines words, speaker attribution, and timestamps, and it accepts custom hotwords for names or domain terms. The documentation lists support for more than 50 languages and code switching without a manual language selector.
The second active path is VibeVoice-Realtime-0.5B, a single-speaker text-to-speech model with streaming-oriented generation and roughly ten minutes of audio context. Microsoft provides a Colab, model weights, a WebSocket demonstration, and file-based inference. It is primarily an English model. Nine other language voices are explicitly experimental, and the guide says those behaviors have not been extensively tested.
The third path is the feature that originally made VibeVoice famous: long-form TTS for conversations lasting up to 90 minutes with as many as four speakers. Its paper, model page, samples, and documentation remain visible, but installation and usage are disabled. Microsoft says it removed the TTS code in September 2025 after discovering uses inconsistent with the research intent. A buyer should not confuse documentation of that model with a supported way to run it.
ASR is the strongest reason to try it
Long recordings are awkward for conventional transcription pipelines. They are often split into chunks, then stitched together with a separate diarization stage. VibeVoice-ASR instead models a long sequence and jointly emits who spoke, when, and what was said. That design is attractive for meetings, interviews, calls, and lectures where speaker continuity across distant sections matters.
The repository includes direct file inference, a Gradio interface, a hosted playground, a Transformers version, a vLLM path, and LoRA fine-tuning instructions. The fine-tuning guide uses prepared audio plus structured transcripts and explains distributed training. This is enough to move from listening to a demo toward a domain experiment without inventing a training pipeline.
There are still practical risks. Open issues report repeated transcript text on some long recordings and a 4-bit MLX decoder falling into a repeated-word loop. Another open report concerns inputs exceeding the model's maximum token sequence. Those reports do not invalidate the architecture, but they are exactly why hour-long claims need testing on your own recording conditions, languages, speaker counts, and noise profiles.
Real-time TTS is narrower than the overview suggests
The 0.5B model is small relative to the rest of the family and has a useful streaming service demo. The guide says a T4 and Mac M4 Pro achieved real-time behavior in Microsoft tests, while warning that weaker devices may need optimization. Network delay can also push audible playback beyond model-side first-chunk timing.
Its product constraints are substantial. It supports one speaker, custom voice prompts are not generally exposed, and English is the intended language. The guide warns about code, formulas, uncommon symbols, very short inputs, and non-speech audio. A TODO still lists streaming text input while audio generation is underway, even though the broader description emphasizes streaming input. That mismatch deserves verification before designing a live assistant around it.
Responsible-use limits also shape the product. The model can produce unexpected, biased, or inaccurate audio, and Microsoft warns about impersonation, fraud, and disinformation. The README recommends disclosure of generated speech and does not recommend commercial or real-world deployment without additional testing. Combined with the removal of the earlier TTS code, that is a firm signal to build consent, disclosure, and abuse controls before any external trial.
Setup assumes a research workstation
Both the ASR and real-time guides recommend NVIDIA PyTorch containers. Their example commands enable GPU access, privileged mode, host networking, shared IPC, and raised memory limits. After cloning, ASR uses an editable install and needs ffmpeg for the Gradio demo. Real-time TTS installs the streamingtts extra, which pins a Transformers version. FlashAttention may need a separate build depending on the container.
This is workable for an ML engineer, but it is not a clean application dependency. The editable-install failure remains an open issue with a documentation fix proposed. Model weights are separate downloads, and the 7B ASR path naturally carries more GPU memory and serving cost than a smaller transcription model. Microsoft now provides VibeVoice-ASR-BitNet and a separate C++ engine for CPU inference, a useful route for edge experiments that lives outside this Python codebase.
Health is excellent, release discipline is light
VibeVoice had more than 52,000 stars at review time, 179 open issues and pull requests combined, and a last push on July 24. Recent merged work added the CPU ASR release and fixed CPU-only behavior, while open contributions address VRAM accumulation, clipping, dependency updates, and demo installation. Activity continued after the last repository push through issue and pull request discussion.
There is no GitHub Release entry, despite the Python package declaring version 1.0.0. That makes commits and model documentation more informative than release tags. Use a pinned revision and a pinned model artifact. VibeVoice earns a place in an ASR evaluation, especially for long, multilingual, speaker-labeled audio. Its TTS options should be judged as separate, constrained research experiments rather than one ready-made speech platform.