The 90-minute TTS feature cannot be installed from this repo
Microsoft's original VibeVoice-TTS promised conversations up to 90 minutes with four speakers, which is still the feature most likely to draw someone to the project. The documentation and samples remain online, but the installation section now says the code is disabled due to widespread misuse. The README dates the removal to September 5, 2025. A buyer cannot treat the old podcast demos as a feature available from the current checkout.
The active family now spans a 7B speech-recognition model, a streaming ASR variant, a separate BitNet CPU project, and a 0.5B realtime TTS model. These share the VibeVoice name and a 7.5 Hz tokenizer idea, yet they solve different jobs and use different runtime paths. Selecting VibeVoice therefore starts with selecting one model. Installing the Python package does not turn every item in the README's model table into a usable service.
The 7B ASR model is the clearest reason to try VibeVoice
VibeVoice-ASR-7B accepts up to 60 minutes of audio in one pass within a 64K token limit. Its output joins recognized words with speaker identity and timestamps, and custom hotwords can bias names or specialist terms. Microsoft documents more than 50 languages and code switching without an explicit language setting. For meetings or interviews where chunk boundaries confuse speaker tracking, that combined output is a concrete reason to evaluate the model.
The 7B path still assumes serious inference equipment. Microsoft's ASR guide recommends an NVIDIA PyTorch container, maps all GPUs into it, and may require a manual FlashAttention install. The Gradio demo adds FFmpeg. Native file inference, Transformers, vLLM, and LoRA fine-tuning give researchers several entry points, but each adds its own compatibility surface. Our 8 GB sandbox verified the package installation and build only. It did not run model weights.
Realtime TTS is a single-speaker English research path
VibeVoice-Realtime-0.5B targets roughly 10 minutes of generated audio from an 8K context window. It uses embedded speaker prompts and supports one speaker, with custom voice requests directed to Microsoft. English is the stated target. Nine additional language voices are offered for experiments, but the guide says their behavior has not been extensively tested. The same guide advertises streaming text input while leaving that function unchecked in its TODO list, a discrepancy worth resolving in your own trial.
Input preparation is part of that trial. The 0.5B guide says code, formulas, and uncommon symbols need normalization, while inputs of three words or fewer may become unstable. It does not generate background noise, music, or sound effects. Microsoft also warns that synthetic speech can enable impersonation and fraud, and it declines to recommend commercial or real-world deployment without further development. That wording places consent, disclosure, and output review on the team using it.
What happened when we ran it
Our run cloned commit 1541f59 into an unprivileged Python 3.12 Bookworm container with 3 CPUs, 8 GB of RAM, and no secrets. Installation succeeded in 87 seconds, adding 141 packages and using 6,146 MB on disk. The package build then succeeded in 9 seconds. Those results show that the source can be packaged in a clean Debian environment, but they do not include any downloaded model weights or GPU runtime.
Tests were skipped because the repository exposed no test script or target. Our scan also found 0 CI workflow files, no Dockerfile, and no tests directory across 110 files and about 20,550 source lines. Pip-audit reported 8 known vulnerabilities in the installed environment. The log supplied no model-quality or inference result, so it would be wrong to turn the successful 9-second build into a claim about transcription accuracy, speech quality, or latency.
A 6,146 MB environment is only the first deployment layer
The checked-out repository measured 149.5 MB, while its installed Python environment grew to 6,146 MB before model weights. That gap matters on developer machines, CI runners, and small containers. The ASR weights, realtime TTS weights, CUDA runtime, and optional FlashAttention sit beyond the successful install we measured. Teams should budget storage and GPU memory for the exact path they choose instead of using the repository build as a proxy for the final service.
VibeVoice provides no Dockerfile in the measured 110-file checkout. The ASR guide's sample container uses privileged mode, host networking, host IPC, unlimited memory locks, and access to every GPU. Those flags are convenient in a research box. A shared service needs a narrower container policy, pinned images and weights, dependency scanning, request isolation, and limits around uploaded audio. The open Gradio issue about one user's Stop action affecting concurrent sessions reinforces that the demo is not a finished multi-user server.
54,529 stars have not produced a tagged release channel
GitHub showed 54,529 stars and 198 combined issues and pull requests on September 28, 2026. The last push was September 3, the same commit our sandbox measured, while new issue and pull-request activity continued on September 28. That is active user interest, not abandonment. Yet GitHub's latest-release endpoint returned no release, and several September fixes remained open pull requests, so adopters must choose and pin source commits themselves.
VibeVoice earns a trial when 60-minute ASR with speaker labels, timestamps, and hotwords could replace several stages in your transcription pipeline. The 87-second install makes that trial possible, while the 6,146 MB environment, missing test target, and 8 audit findings set the terms. Pick one model path, isolate it, and test the recordings your users will actually send. The disabled four-speaker TTS demo is history, not a procurement plan.