AivisSpeech: a VOICEVOX-based Japanese TTS editor built around AIVMX models

What AivisSpeech is, and who it is actually for

AivisSpeech is a Japanese speech synthesis application built on the editor UI of VOICEVOX, with the AivisSpeech Engine embedded so that voice generation happens without a separate server the user has to manage. The repository describes it as an AI Voice Imitation System, and the primary language is TypeScript, with Vue, Electron and Vite in the topic list. The engine itself lives in a separate repository, Aivis-Project/AivisSpeech-Engine.

The audience is narrower than the name suggests. The README states that the Engine supports Japanese speech synthesis only, in the same way as VOICEVOX ENGINE. AIVM metadata can describe multilingual speakers, but the Engine will not synthesise English or Chinese even when the underlying model was trained for it. If you need English output, this project is the wrong layer.

What you get instead is a desktop application with a graphical editor, model management from the UI, and a bundled engine. The README points end users at the official site and at three documents under public/ (howtouse.md, qAndA.md, contact.md) rather than at the repository README, which is written for developers.

How the AivisSpeech Engine and AIVMX files fit together

The data flow is straightforward. The editor talks to the AivisSpeech Engine over HTTP on 127.0.0.1 port 10101, which is the host value shown in the VITE_DEFAULT_ENGINE_INFOS example in the README. The Engine loads speech synthesis models from a per-OS Models directory and exposes an API described by openapi.json in the repository root.

The model format is the interesting part. AIVM is an open file format that bundles a trained model, hyperparameters, style vectors and speaker metadata (name, description, licence, icon, voice samples) into one file. An AIVM file is Safetensors with AIVM metadata added; an AIVMX file is ONNX with the same metadata added. The Engine is a reference implementation of the AIVM specification but deliberately supports only AIVMX. The README gives the reason: dropping PyTorch shrinks the install size and lets ONNX Runtime handle CPU inference.

That choice has a cost. AIVM files, which are the more portable form for tooling that works with PyTorch, cannot be loaded directly. You convert first. The README points at AIVM Generator, a web tool for producing AIVM or AIVMX files from existing models and for editing metadata on existing files.

Supported model architectures are Style-Bert-VITS2 and Style-Bert-VITS2 (JP-Extra). Nothing else is listed.

Installing AivisSpeech and generating your first line of Japanese speech

End users do not build this from source. The README directs them to the download page at aivis-project.com/speech, and the Engine has its own releases page. The UI can add speech synthesis models, and the README recommends that route for end users over placing files by hand.

If you do place files manually, the location depends on the OS. The README lists these three paths, and notes that the Engine prints the real one at startup as Models directory:.

# Windows
C:\Users\(username)\AppData\Roaming\AivisSpeech-Engine\Models

# macOS
~/Library/Application Support/AivisSpeech-Engine/Models

# Linux
~/.local/share/AivisSpeech-Engine/Models

Developers building the editor from source need Node.js 22.11.0, per the README, and the steps differ from upstream VOICEVOX. The dependency install and environment file copy are the first two commands.

npm ci
cp .env.development .env

The README notes that the copied .env does not need editing. On macOS, .env.production does: the executionFilePath value must change from "AivisSpeech-Engine/run.exe" to "../Resources/AivisSpeech-Engine/run". That setting is used when launching a product build created by npm run electron:build.

The Engine has to be running before the editor. Its development environment is built separately, and the README shows the command as poetry run task serve inside the AivisSpeech-Engine directory. Once both are up, npm run electron:serve starts the Electron build of the editor, and npm run browser:serve starts the browser build. Note that package.json pins node to ">=22.14.0 <23" and declares [email protected] as the package manager, while the README's setup instructions use npm ci; that mismatch is worth knowing before you script a CI job.

Where AivisSpeech breaks, and where it is the wrong tool

The development policy section is unusually blunt, and it describes the maintenance model rather than a bug. The project keeps changes to VOICEVOX minimal so that tracking upstream stays cheap. Refactoring is not done, because conflicts with VOICEVOX are expected and the maintainers state they are not deeply familiar with the whole codebase. Features AivisSpeech does not use, such as singing synthesis, are not deleted; they are commented out. Documentation is not updated, and the README says plainly that the documents under docs/ are inherited from VOICEVOX without modification and that their contents may not apply. Test code is not updated either, and the README states that E2E tests in particular do not work properly because the UI has changed substantially.

Practically, that means you should not expect the docs/ tree to describe AivisSpeech behaviour, and you should not treat the test suite as a signal that a given path works. It also means upstream VOICEVOX changes arrive quickly but are not re-verified here.

Platform support is another boundary. Only Windows and macOS are listed. Windows 10 must be 22H2 or later; the README warns that older and LTSC builds of Windows 10 have been reported to crash the Engine on startup. On macOS, Intel machines are not actively verified, and the README recommends Apple Silicon. The application needs at least 1.5GB of free RAM.

Finally, the language limit is a hard stop for some users. Even a model with multilingual speakers defined in its AIVM metadata will only produce Japanese through this Engine.

AivisSpeech compared with VOICEVOX and with hosted TTS APIs

The closest comparison is VOICEVOX itself, and the README frames the relationship directly: AivisSpeech is based on the VOICEVOX editor UI and follows the latest VOICEVOX release with minimal modification. The difference is the engine and the model format. VOICEVOX is the upstream project; AivisSpeech substitutes AivisSpeech Engine and the AIVM/AIVMX model format, and adds UI for adding models. If you already run VOICEVOX, the reason to switch is the AIVMX model ecosystem, not the editor, which is largely the same code.

Hosted APIs such as ElevenLabs or CoeFont are a different trade entirely. They require no local model files, no 1.5GB of free RAM and no ONNX Runtime, and they are not limited to Japanese. In exchange, synthesis happens on someone else's infrastructure, per-use costs apply, and the voice set is whatever the provider offers. AivisSpeech runs the model on your own CPU and lets you supply any AIVMX file you have the rights to use. The README does not compare AivisSpeech to these services, so treat the choice as one of deployment model rather than quality.

For Japanese-only, offline, self-hosted synthesis with Style-Bert-VITS2 voices, AivisSpeech is a reasonable fit. For multilingual output or zero local setup, it is not.

Licence and the real cost of keeping this fork current

AivisSpeech inherits only the LGPL-3.0 half of VOICEVOX's dual licence, and the repository states this explicitly. The LICENSE file at the root carries the identifier. LGPL-3.0 matters if you redistribute a modified build: the usual obligations around source availability and relinking apply. This is a description of what the repository says, not legal advice; check the licence text and your own distribution model.

Model files are a separate question. AIVM metadata includes a licence field for each speaker, so the terms attached to a voice come from the model, not from AivisSpeech. The README does not state what licence the bundled or default voices carry, so verify that per model before publishing audio.

Upgrade cost is the more interesting number. Because the project avoids refactoring and keeps edits minimal, rebasing onto a new VOICEVOX release should be small. But the README also says documentation and tests are not maintained, so the verification work after each rebase falls on you. Recent releases listed for the repository are development and preview builds (1.1.0-dev, 1.1.0-preview.3, 1.1.0-preview.4), and the last push to the repository was on 2026-07-12. There is no stable 1.1.0 release listed, so anyone deploying this should decide whether a preview build is acceptable.

Editorial conclusion

AivisSpeech suits Japanese-speaking users on Windows 10 22H2 or later, Windows 11, or macOS 13 Ventura or later who want a desktop editor and are happy to run Style-Bert-VITS2 models in AIVMX form; the README states the Engine supports Japanese synthesis only, so anyone needing English or Chinese output should look elsewhere. Before committing, confirm that your machine has at least 1.5GB of free RAM, that your CPU is Apple Silicon if you are on a Mac, and that the AIVMX models you want actually exist, since the Engine deliberately refuses plain AIVM files and depends on ONNX Runtime for CPU inference.

Frequently asked questions

Official sources

1. Aivis-Project/AivisSpeech on GitHub
2. License: LGPL-3.0
3. Project website
4. README
5. Releases