OpenVINO AI Plugins for Audacity: local neural audio tools in an effects chain
A set of AI-enabled effects, generators, and analyzers for Audacity®.
At a glance
- What is it?
- Five neural effects for Audacity that run entirely on your own CPU, GPU or NPU, covering stem separation, noise suppression, music generation, transcription and upscaling.
- Who is it for?
- The strongest reason to use these plugins is architectural rather than model quality: everything runs locally, so stems, transcriptions and voice recordings never leave the machine, which is rarely true of web based audio tools. The cost is that each effect is pinned to a specific Audacity version, and the version numbers in the release tags read like a compatibility contract rather than a product roadmap.
- Can I use it commercially?
- Yes, with conditions. GPL-3.0 is a copyleft licence: if you distribute software that includes it, you must release that software's source code under the same licence. Running it internally without distributing it does not trigger that obligation.
- Is it still maintained?
- Yes. The repository last received commits 26 days ago.
- What is it written in?
- Mainly C++, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 20, 2026, and from our analysis. They are not legal advice.
Editorial analysis
Five effects, all of them running on your own hardware
The pitch is stated in one sentence at the top of the README: these are AI enabled effects, generators and analyzers for Audacity, and they run entirely on your PC with no internet connection needed. OpenVINO is the runtime underneath, and its role is to dispatch models onto whichever accelerator the machine has, whether that is a CPU, a GPU or an NPU.
That local execution is the feature that matters most for audio work. Uploading a voice recording to a web service to clean it up means handing over the one recording you cannot afford to leak. Running the same model through a local plugin keeps the file on the machine and reduces the round trip to nothing.
The feature set covers five distinct jobs. Music Separation splits a mono or stereo track into individual stems, specifically drums, bass, vocals and other instruments. Noise Suppression removes background noise from an audio sample. Music Generation and Continuation uses the MusicGen LLM to produce snippets of music or to continue an existing passage. Whisper Transcription generates a label track containing a transcription or a translation for a selected span of speech or vocals. Super Resolution upscales and enriches audio for clarity and detail.
Each of those has its own documentation file under `doc/feature_doc/`, with a directory named after the feature, so the README acts as an index into detailed per effect pages rather than as the manual itself.
Where the models come from, and what got ported to C++
The acknowledgements section is unusually detailed, and reading it tells you exactly what this project is: a set of C++ ports and OpenVINO conversions of existing research models, rather than a new set of models.
Music Generation and Continuation use Meta's MusicGen, with support currently for MusicGen-Small and MusicGen-Small-Stereo hosted on Hugging Face. The README states the text to music pipelines were ported from Python to C++, referencing logic from the Hugging Face transformers project. Music Separation uses Meta's Demucs v4 model, which has been ported to work with OpenVINO.
Noise Suppression draws on two sources. One is a dense U-Net model from the OpenVINO Open Model Zoo. The others are DeepFilterNet2 and DeepFilterNet3, ported from a Rust implementation, with the README crediting both the original project and a contributor's fork that helped clarify the Rust code, noting that some of the C++ implementation is based on an offline script from that fork. Two academic papers are cited for that work, one on real time speech enhancement on embedded devices and one on perceptually motivated real time speech enhancement.
Super Resolution is a port of the AudioSR project, taken from Python and PyTorch to C++ and OpenVINO IR, with its paper cited. Super Resolution also carries a citation from the AudioSR paper itself.
Whisper Transcription is different in kind, since it uses whisper.cpp with an OpenVINO backend rather than a converted model. So the plugin set is a mix of three porting strategies, which is worth knowing when a particular effect misbehaves on your hardware.
Version tags are a compatibility contract, not a roadmap
The release tags follow the pattern `v3.7.1-R4.2`, and reading them makes the constraint obvious. The leading numbers are the Audacity version the plugins are built against, and the R number is the plugin revision. Each release says so in a warning: this plugin release is only compatible with that specific 64-bit Audacity release for Windows.
The most recent release is v3.7.1-R4.2, published in December 2024. It added Audio Super Resolution as a new feature and updated OpenVINO to 2024.6.0. Its notes also carry a useful piece of advice: the same plugins are compatible with Audacity 3.7.3, so anyone on 3.7.1 or 3.7.2 can simply install the newer Audacity rather than hunting for a matching plugin build.
The release before it, v3.7.1-R4.1, updated Audacity compatibility from 3.7.0 to 3.7.1 and moved OpenVINO to 2024.5.0. The README of that release is candid about a real bug: earlier versions had an incompatibility on certain machines that required users to manually delete an NPU plugin library file, and the OpenVINO update removed the need for that workaround. That is a useful signal about the failure mode to expect, which is a plugin that will not initialize rather than one that produces wrong audio.
Version v3.7.0-R4.0 shows the other class of fix, a problem with Windows usernames containing special characters, tracked as issues 313 and 293. The last push on the default branch was 2026-09-11, well after the newest tag, so there is work in the repository that has not yet been packaged into a release.
Installation, the model download trap, and Linux
Installation is not described in the README at all, which is unusual but reasonable for something whose packages are version specific. The README simply sends you to the releases page for the installation packages and instructions for the latest Windows release.
Two warnings from the release notes are worth reading before installing rather than after. The first is that if you have multiple Audacity versions installed, the installer may select the wrong location by default, so you should check the selected path and adjust it with the browse button if needed. The second is the one that catches people out: if you have previously installed a 3.7.X, 3.6.X or 3.5.X plugin release, the recommendation is not to uninstall the previous version but to install the new one and select the option to install no models from the dropdown, so you do not have to download every model again.
That second point only makes sense because the models are large and are fetched separately from the plugin itself, which is what makes the local-first design practical to distribute.
Windows is the only platform with a prebuilt installer, but it is not the only platform you can build. The repository has separate build instructions for Windows and for Linux under the `doc/build_doc/` directory, and the v3.7.0 release notes note that while only a Windows installer is provided, the project can be compiled elsewhere. The tree is small and gives a hint about the structure, with a `mod-openvino/` directory for the OpenVINO integration, a `tools/` directory, and the documentation tree.
Licensing, credit, and where the project sits
The plugins are released under the GNU General Public License version 3, with a badge in the README and a `LICENSE.txt` in the repository root. That is a meaningful choice for an Audacity add on, and it is worth understanding before you build on this: the plugin code is copyleft, which is the same license Audacity itself uses.
The upstream models carry their own terms separately, and the README credits each source rather than treating them as interchangeable. Meta's Demucs and MusicGen, the OpenVINO Open Model Zoo dense U-Net, DeepFilterNet, and AudioSR are each attributed to their authors, and whisper.cpp is credited as the transcription backend. For the OpenVINO notebooks, the README is direct about the relationship: the project learned from that collection of Python notebooks and is still using it to follow current practices for AI pipelines.
The Audacity development team and Muse Group are thanked for their support, which matters because a version locked plugin set needs upstream cooperation to keep working. The project also carries a code of conduct, a contributing guide and a security policy.
At 2,119 stars, 135 forks and 56 open issues, this is a substantial project with a real support load, and the repository welcomes pull requests of any size. The open issue count against a release cadence that produces roughly one build per Audacity version is the practical constraint to weigh: the plugins are good, and they are tied to an Audacity version you do not control.
Editorial conclusion
The strongest reason to use these plugins is architectural rather than model quality: everything runs locally, so stems, transcriptions and voice recordings never leave the machine, which is rarely true of web based audio tools. The cost is that each effect is pinned to a specific Audacity version, and the version numbers in the release tags read like a compatibility contract rather than a product roadmap. Install the Windows package for the Audacity build you actually have, choose to skip model downloads when upgrading, and read the compatibility warning in the release notes before assuming a newer Audacity will work. The feature documentation lives per effect under `doc/feature_doc/`, which is where the model sources are also credited.
Frequently asked questions
How do I install the OpenVINO AI plugins for Audacity?
From the repository's releases page, using the Windows package that matches your Audacity version. The release tags encode the requirement, for example v3.7.1-R4.2 targets Audacity 3.7.1 and is also compatible with 3.7.3. When upgrading, the release notes recommend installing over the previous version and choosing the option to install no models so you avoid downloading everything again.
What AI features do the OpenVINO Audacity plugins add?
Five. Music Separation splits a track into drums, bass, vocals and other stems. Noise Suppression cleans background noise. Music Generation and Continuation uses the MusicGen model to create or extend music. Whisper Transcription produces a label track with a transcription or translation. Super Resolution upscales audio for clarity and detail.
Is OpenVINO AI free?
The plugins in this repository are released under the GNU General Public License version 3, and they run entirely on your own machine with no internet connection required, so there is nothing to pay per use. OpenVINO itself is a separate Intel project with its own licensing, and the upstream models credited in the README carry their own terms as well.
Does OpenVINO for Audacity need an internet connection?
Not to run. The README states the AI features run locally on your PC with no internet connection necessary, and OpenVINO is used to execute the models on whatever accelerator the system has, such as a CPU, GPU or NPU. Models are downloaded separately from the plugin, which is why upgrading lets you skip the model download step.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/intel-openvino-plugins-ai-audacity)