Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

Analysis · September 4, 2026 · repo

VoiceStudiolocal-aiprivacyagent-infrastructureregulation

VoiceStudio's Badge Says AGPL-3.0. Its Default Model Weights Are CC-BY-NC.

The repo license covers the code. The model card covers the output. Only one of them shows up in GitHub's sidebar.

There is a note at the bottom of VoiceStudio's license file explaining why the file exists at all. LICENSE has to contain the verbatim AGPL-3.0 text and nothing else, the project writes, so that GitHub's license detection and "the corporate license scanners that gate adoption" identify it as AGPL-3.0-only rather than falling back to NOASSERTION. Everything the project actually needs you to understand about what you may do with its output had to be moved into a second file, because the first one is written for a machine.

That machine is the reason most people will never learn that VoiceStudio's default downloaded model weights are CC-BY-NC. Non-commercial. It is in the notice, stated plainly. It is nowhere near the badge.

What the notice actually says

VoiceStudio is a fully local voice platform covering cloning, voice design, dubbing, dictation, transcription and audiobook creation, in a claimed 646 languages, with no account and no API key. Version 0.5.1 shipped on August 28, 2026. It was previously called OmniVoice Studio, and it trended on GitHub this week in a run otherwise dominated by agent control-plane tools.

Its LICENSE-NOTICE.md is one of the more careful licensing documents I have read on an AI repository, and it is worth quoting rather than summarizing. On the application itself:

VoiceStudio is free and open-source software, licensed under the GNU Affero General Public License, Version 3 (AGPL-3.0). You are free to use, copy, modify, and redistribute it. That includes commercial and internal business use of the application itself. Model weights, tokenizers, and other third-party assets retain their own terms; this application license does not grant or summarize rights under those separate terms.

Then, further down, under Scope:

Downloaded model weights are not relicensed by VoiceStudio. The default k2-fsa/OmniVoice model card identifies its code as Apache-2.0 and pretrained weights as CC-BY-NC. Its audio_tokenizer/LICENSE contains separate Boson Higgs Audio 2 and Meta Llama community terms. A commercial license for VoiceStudio-owned code does not replace any of those terms.

Read those two passages together and the picture is unambiguous. The app is AGPL-3.0 and you may use it commercially. The weights it downloads on first run, in the default configuration, are CC-BY-NC. The audio tokenizer sits under two more sets of terms that are neither of those. And the commercial license the project sells, for organizations that want to embed VoiceStudio without the AGPL copyleft obligation, explicitly does not touch any of it. Pricing on that license is listed as coming soon.

None of that is hidden. All of it is invisible to a license scanner reading the repository root.

The structural problem, not the project's problem

VoiceStudio is the well-behaved example here. It wrote the second file. Plenty of projects distributing model downloads do not, and the absence of a notice is not evidence that the weights are permissive; it is evidence that nobody wrote it down.

The mechanism is worth naming precisely. A repository license is a static document that ships with the clone. Model weights are a runtime dependency fetched over the network after install, from a different host, under a different license, chosen by a config value. Your compliance posture is therefore determined by a file you did not clone, selected by a setting you probably did not change, described on a model card you have never opened.

The bundled omnivoice/ Python package is Apache-2.0 by its upstream author, Han Zhu, and is "not relicensed here." That is a code dependency and it behaves like one. The weights that same package pulls behave like nothing in the traditional dependency model at all, and CC-BY-NC on those weights is not an edge case: it is the default path for anyone who installs and runs.

Why the timing sharpens this

Microsoft AI released MAI-Transcribe-2 on September 3 at ten cents per hour of audio, a limited-time price through the end of 2026. That is the comparison a team actually faces, and price is the least interesting axis of it.

A hosted API gives you one counterparty, one commercial agreement, and one company answerable for the terms. Local gives you privacy, offline operation, no metered bill, and full responsibility for a license stack you assembled by clicking install. For internal work the local option wins on almost every axis and the licensing question is close to moot. The moment the output becomes a product, the axes swap, and the team that never opened the model card is the one that finds out late.

Put this into practice

Twenty minutes of work, done once per tool, closes most of this.

Split the question in two before you install. Question one: may I run, modify and distribute this code, and what do I owe if I offer it over a network? Question two: may I sell, publish or commercially distribute what it produces? For VoiceStudio the notice answers the first (yes, including commercial and internal business use) and points you elsewhere for the second. Write both answers down. They are different answers from different documents.

Find the second file before you find the badge. Look for LICENSE-NOTICE.md, NOTICE, THIRD-PARTY, or a licensing section in the docs. If a project ships model downloads and has none of those, that silence is your finding, and the correct next step is reading the model cards yourself rather than assuming.

Run it once in a scratch environment and inventory what landed. The point of the first run is the download. Install with the project's own command:

curl -fsSL https://voicestudio.sh/install | sh

Or use the headless build, which keeps the experiment contained. The current Quick Start on Docker Hub uses the :latest tag and requires an API key environment variable, so the full shape is:

export OMNIVOICE_API_KEY=...
docker run -d --name omnivoice \
  -p 127.0.0.1:3900:3900 \
  -e OMNIVOICE_API_KEY="$OMNIVOICE_API_KEY" \
  -v omnivoice-data:/app/omnivoice_data \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  palashdeb/omnivoice-studio:latest

Then look in ~/.cache/huggingface and open the model card for every weight file you find. The license you are operating under is the union of what is in that directory plus the repository license, and the repository license is usually the permissive half.

Check whether a permissive alternative engine exists before you commit. VoiceStudio's architecture is pluggable and credits a long upstream list including WhisperX, Demucs, Pyannote, CTranslate2, AudioSeal, Sherpa-ONNX, GPT-SoVITS and PocketTTS. If the default weights are CC-BY-NC and your use is commercial, the question is not "can I use this tool" but "which engine can I configure it to use." That is a solvable problem and it is much cheaper to solve on day one.

Pin the engine, and say why in a comment. When you settle on a configuration because its terms permit your use, pin it and annotate the pin as a legal constraint rather than a technical one. The next person upgrading for a quality win should hit that note before they hit the merge button.

Honest limitations

I read VoiceStudio's LICENSE-NOTICE.md in full and I have quoted it directly rather than paraphrasing, because a paraphrase is what got me into trouble on the first draft of this piece. Several secondary summaries of this project circulating right now assert a sentence about "downloaded weights and audio tokenizer terms" that is close to the notice's meaning but is not its actual language, and I nearly published one of them. That is a small live demonstration of the thesis: the licensing facts about AI repositories degrade fast once they leave the primary document.

What I have not done is open the k2-fsa/OmniVoice model card myself, or read the audio_tokenizer/LICENSE that carries the Boson Higgs Audio 2 and Meta Llama community terms. My claim about CC-BY-NC is VoiceStudio's characterization of that card, which is a primary source about a secondary document. If your use is commercial and this matters to your decision, read the card. I am not a lawyer and this is not legal advice; it is a map of where the documents live.

The star count is not quotable this week. The project's own site showed 11,000 at read time while GitHub's rendered pages served 10k, 14.9k and 17.6k across three loads in one session on September 4, which is stale-cache behavior rather than growth. Adoption proxies that do not depend on that number are healthier: Docker Hub shows a bucketed 50K+ pulls and GitHub Container Registry shows about 19.3K downloads.

There are rough edges unrelated to licensing. The published macOS DMG is ad-hoc signed and not notarized, so Gatekeeper will object and you should verify the SHA-256 the project publishes on its download page. The macOS Intel build requires a remote backend because the local Python backend is unavailable on that architecture, which puts an asterisk on "fully local" for Intel Macs. Hardware guidance is around 8 GB of RAM with 16 GB or more recommended, roughly 10 GB of disk for models, and an optional GPU, with 4 GB of VRAM workable using CPU offload.

And the counterargument stands: for someone running this on their own laptop to narrate their own book, none of the above changes anything. Internal and personal use is the easy case and it is most of the use. Everything here is scoped to the moment the audio leaves the building.

What to do with this

Take the local AI tool you are most likely to ship something on. Note its license badge. Then go find the document that describes what the weights it downloads are licensed under, and time how long it takes.

If it takes five minutes, the project did its job. If you cannot find the document at all, you have learned the most useful thing available about that project today. And if you find something like VoiceStudio's notice, saying the app is yours to use commercially while the default pretrained weights are CC-BY-NC, then you know exactly which question is still open and exactly which model card answers it.

GitHub's license detector will keep telling you AGPL-3.0-only, and it will keep being right about the code and silent about everything else. The projects that write the second file are doing you a favor. Read it.

Sources: VoiceStudio LICENSE-NOTICE.md · VoiceStudio on GitHub · VoiceStudio v0.5.1 release (28 August 2026) · voicestudio.sh download and verification page · Microsoft AI, MAI-Transcribe-2 (3 September 2026)

Medium metadata

Title: VoiceStudio's Badge Says AGPL-3.0. Its Default Model Weights Are CC-BY-NC.

Subtitle: The repo license covers the code. The model card covers the output. Only one of them shows up in GitHub's sidebar.

Tags: Open Source, AI, Licensing, Machine Learning, Software Development

Recommended publication: Tkay Nation's Writes

Canonical: import from the fervorai.dev URL

Kicker for social: VoiceStudio moved its real licensing explanation into a second file because LICENSE has to stay machine-readable. The thing in that second file: the default weights are CC-BY-NC.