Vocello: A Local, Private Voice Studio for Apple Silicon

Cloud voice tools are convenient, but they change the shape of the work: your script leaves your computer, model access depends on a service, and a private recording becomes another upload to reason about. Vocello takes the opposite route. It is a local voice studio for Apple Silicon, built to generate speech on the machine after its model packages have been installed.

The current Mac release is Vocello 2.4.0 for macOS 26 and Apple Silicon, while the iPhone app is a public TestFlight beta. The project is native Swift plus MLX rather than a Python wrapper around a local server, so this article separates the ready-to-install Mac release from the source-build path for developers.

What It Does

Built-in Voice gives you nine Qwen3 speakers for the quickest path from text to audio. Vocello detects the language automatically, or lets you choose it manually, and exposes delivery controls so a sentence can be made more expressive, balanced or consistent instead of sounding like one fixed preset.

Voice Design is the more interesting mode when you want a character rather than a named speaker. Describe the voice in ordinary language — its age, timbre, accent, energy or delivery — then save a design that can be reused. That makes it practical for narration, prototypes and recurring characters without starting from a sample every time.

Voice Cloning accepts a reference recording that you own or are permitted to use, and the app keeps a visible consent acknowledgement in the flow. Vocello supports ten languages, turns scripts over 900 characters into long-form projects, streams their segments in order, and keeps the segment map in History so a long take remains editable.

How to Install It

There is no App Store release for the Mac app at the time of writing. Use the official Vocello 2.4.0 Release, or download its direct Mac DMG. You need macOS 26 or later on Apple Silicon and at least 8GB of memory; 16GB or more is recommended for the Quality model and long projects. The iPhone build is separate and available through the project's public TestFlight beta.

# Download the signed and notarized Mac release:
open https://github.com/PowerBeef/Vocello/releases/download/v2.4.0/Vocello-macos26.dmg

# Open the DMG and drag Vocello.app to /Applications.

The Mac DMG is Developer ID signed, notarized and stapled. After copying the app, open Settings and install the model packages: the recommended Speed packages use about 6GB, while adding Quality can bring the total close to 15GB. The first model download needs an internet connection and comes from Hugging Face; once the packages are present, speech generation, scripts, references, saved voices, history and generated audio stay local unless you export them. Updating a 2.x installation preserves the app data and installed models.

How to Use It

Start with a short script and the Built-in Voice mode. Choose a speaker, select the language if automatic detection is not what you want, adjust the delivery style, and generate. The result can be listened to, saved in History, and exported as audio. A per-request seed is recorded in History too, so a fixed seed can reproduce a take when the text and settings are unchanged.

Switch to Voice Design when the voice itself is part of the brief. Use Voice Cloning only for a reference you own or have permission to use; the app's consent acknowledgement is a safety boundary, not a replacement for that permission. On the Mac, reference clips can be imported in common formats including WAV, MP3, AIFF, M4A, FLAC, OGG and WebM.

# After installing Vocello:
# Settings → Model downloads → install Speed packages
# Then choose Built-in Voice, Voice Design, or Voice Cloning.

For a long narration, paste the full script rather than trying to manage every chunk yourself. Vocello turns a script above 900 characters into a project, generates the segments one after another, and joins them into a finished file. The iPhone 2.4.0 beta also supports regenerating one segment, which is a useful correction loop for a single sentence without rerunning the whole project.

# Build the developer CLI only when you need it:
./scripts/build.sh cli
build/vocello modes
build/vocello speakers list

Voice Workflows and Output

Vocello is aimed at the space between a one-line demo and a full production pipeline. You can draft a voice-over, create a recurring character, generate a multilingual prototype, or turn a long script into a sequence that remains understandable in History. The supported languages are Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.

Under the Hood

The repository's core is VocelloQwen3Core, a first-party Swift runtime on MLX derived from mlx-audio-swift. The app does not require Python, a separately managed local server or bundled model weights. One engine is shared across the Mac app, iPhone app and CLI path, while model packages are downloaded separately so the application bundle does not have to carry them.

Long-form generation is streamed in bounded frames instead of holding every intermediate buffer forever. In the repository's own measurement, a comparable take peaked around 8GB and later stayed near 3GB, while a twelve-segment project finished close to its starting footprint. The engine produces 24kHz mono 16-bit WAV output, and the repository tracks real-time-factor benchmarks rather than promising one speed for every Mac.

Repository and Releases

The source lives in the PowerBeef/Vocello repository, and the current Mac download is the v2.4.0 release. The repository is MIT-licensed and includes the app, the shared engine, build scripts, documentation and release metadata. The Mac DMG is the normal user path; you do not need Xcode just to install it.

The CLI is not included in the Mac DMG. Build it from source only when you need command-line generation or want to contribute: the repository scripts expect full Xcode 26 on macOS 26 and Apple Silicon, and Command Line Tools alone are not enough. If you still need macOS 15 support, the project points to the legacy QwenVoice 1.2.3 release rather than pretending Vocello 2.x has that compatibility.

What to Verify Before Installing

Check four things before downloading: your Mac is Apple Silicon, it runs macOS 26 or later, it has enough free disk space for the model packages, and you are comfortable with a one-time Hugging Face download. If you record references, allow microphone access; if you use transcript auto-fill, macOS may request Speech Recognition permission for its on-device assets. The iPhone TestFlight beta is a separate experiment, not a substitute for the Mac release.

Where It Falls Short

Vocello 2.x is intentionally narrow: Apple Silicon only, macOS 26 or later, and no current Intel or older-macOS build. The iPhone app is still a public beta, and the Mac and iPhone model sets are not identical. If you need a small, instant download or a browser workflow, a local model studio will feel heavier than a hosted TTS page.

Model downloads are a real part of the setup, and the Quality packages consume more memory and disk than Speed. Voice Cloning also depends on the reference recording and the selected model; it is not a promise that every voice will reproduce cleanly. Finally, local generation is not permission to use another person's identity: you still need authorization for a reference and should disclose AI-generated audio when publishing a cloned real voice.

Vocello adds an inaudible AudioSeal watermark and a machine-readable provenance note to WAV output. Its documentation also notes that disclosure obligations may apply in some jurisdictions, including the EU. Treat that as project guidance to check for your use case, not as a substitute for legal advice.

License

The application source is released under the MIT License. In practical terms, you can inspect, modify, use and redistribute the code, including in commercial software, as long as the copyright and licence notice remain with the relevant distribution. The licence comes without a warranty.

That permission applies to the project code, not automatically to every model, dependency, voice reference or generated work. Qwen3-TTS, MLX, mlx-audio-swift, downloaded weights and bundled assets have their own terms, and a source licence does not grant rights to clone a real person's voice. Read the repository notices before redistributing a modified app or model package.

Should You Install It

If you have an Apple Silicon Mac on macOS 26 and want a private voice workflow that can stay on-device, start with the signed Vocello 2.4.0 DMG. Use Built-in Voice for a quick result, Voice Design for a repeatable character, and Voice Cloning only when the permission and disclosure story is clear. Build from source only for the CLI or development work; try the iPhone TestFlight beta separately if you want the mobile companion.

# Download the signed and notarized Mac release:
open https://github.com/PowerBeef/Vocello/releases/download/v2.4.0/Vocello-macos26.dmg

# Open the DMG and drag Vocello.app to /Applications.