A local, private voice studio for Apple Silicon — on Mac and iPhone. Write a script, shape how the voice should sound, and generate speech right on your device.
On your Mac today · iPhone arriving soon.

Formerly QwenVoice — now Vocello 2.0.

• 🎙️ Three ways to make a voice — pick a built-in speaker, describe one in plain language, or clone from a reference clip you have rights to (record it right in the app, or import a file).
• 🔒 Private by default — after a one-time model download, every line renders on your device. No scripts uploaded, no audio sent to a cloud TTS service.
• ⚡ Native Swift + MLX — no Python runtime, no bundled weights, no per-line meter and no cloud queue.
• 📱 iPhone arriving soon — the same on-device engine on Apple Silicon iPhone; on-device generation already works, with the App Store / TestFlight distribution lane still in progress.

Vocello is coming to iPhone — the same local, private engine, running fully on-device on Apple Silicon. Write a script, pick or describe a voice, and generate speech without a cloud round-trip, exactly like the Mac app. On-device generation already works. The remaining piece is the App Store / TestFlight distribution lane (not GitHub Releases), which is still in progress. No public release date yet. This is what the native Swift + MLX rebuild was for: replacing the old bundled Python runtime with an engine that runs entirely on-device — the only way to bring Vocello to iPhone. Want to follow along? Star ⭐ and watch 👀 the repo for updates.

• Private by default. After models are installed, generation runs locally and your scripts, history, and generated audio stay in local app storage unless you export them.
• No subscription meter. Download the models you want, then generate on your own hardware without paying per line or waiting on a cloud queue.
• Three voice workflows. Use a built-in speaker, describe a new voice, or clone from a reference clip you own or have permission to use — recorded in the app or imported.
• Built for Apple Silicon. A native Swift + MLX engine (replacing the old bundled Python runtime) keeps generation local, private, and fully on-device — and it's what makes the iPhone app possible (Python can't ship on iPhone; on-device MLX can).

Custom Voice Pick one of nine built-in Qwen3 speaker presets, set delivery, and generate a clean spoken line. The fastest path when you want a consistent voice right away.	Voice Design Describe the voice you want in plain language, then write the script. Vocello shapes the take from that brief.
Voice Cloning Record a short reference clip with your Mac's microphone, or import one (WAV, MP3, AIFF, M4A, FLAC, or OGG). The optional transcript can auto-fill with on-device transcription. Only clone voices you own or have permission to use.	Model downloads Install and manage Speed and Quality packages for each voice mode from Settings. Generation screens own the Speed/Quality choice while you write.

1. Download Vocello-macos26.dmg.
2. Open the DMG and drag Vocello.app to /Applications.
3. Open Vocello.
4. Go to Settings → Model downloads and install the voice models you want.
5. Generate from Custom Voice, Voice Design, or Voice Cloning.

No Python setup or local server is required — install the app, download models from Settings, and generate locally.

The DMG is signed with an Apple Developer ID certificate and notarized with a stapled ticket, so the first launch opens with a double-click (no right-click bypass). To verify:

xcrun stapler validate Vocello-macos26.dmg              # "The validate action worked!"
spctl --assess --type install -vv Vocello-macos26.dmg   # accepted, source=Notarized Developer ID

A release-metadata.txt (commit SHA, Xcode version, SDK, marketing version, build number) is attached to the same release for build provenance.

• macOS 26.0+ on an Apple Silicon Mac — available now.
• iPhone (iOS 26.0+) on Apple Silicon — arriving soon via App Store / TestFlight.
• Voice models installed from Settings → Model downloads.

Speed models are smaller 4-bit packages for faster startup and lower memory use. Quality models are larger 8-bit packages for devices with more headroom.

Vocello 2.0.0 is the current stable macOS release. For macOS 15, use QwenVoice v1.2.3; no 2.x backport is planned. Every macOS GitHub Release ships a notarized, stapled, Developer ID–signed DMG — a normal double-click install with no Gatekeeper workarounds.

• Speech generation runs locally after models are installed.
• Generated audio, recorded reference clips, and history stay in local app storage unless you export them.
• Model downloads come from Hugging Face when you install a voice model.
• Recording a reference clip and transcript auto-fill ask for the Microphone and Speech Recognition permissions on first use. Both run entirely on your Mac — recognition is on-device only, and nothing is sent to Apple or anyone else. (Transcript auto-fill additionally requires Siri to be enabled, a macOS requirement; the app explains this and links the right Settings pane.)
• Voice cloning should only be used with voices you own or have permission to use.

The main branch contains the current Vocello codebase (macOS app, iPhone app, and the vocello CLI). The stable macOS release is tagged v2.0.0.

Vocello's engine is native Swift + MLX — no Python, no bundled weights. On macOS it runs out-of-process in an isolated XPC service; on iPhone it runs in-process, fully on-device. Architecture, engine invariants, and release policy live in CLAUDE.md.

git clone https://github.com/PowerBeef/QwenVoice.git
cd QwenVoice
./scripts/regenerate_project.sh
open QwenVoice.xcodeproj

The Xcode project is generated from project.yml (edit it, not the .xcodeproj, then rerun regenerate_project.sh). SPM dependencies — MLX, Swift HuggingFace, GRDB, and the vendored mlx-audio — are deliberately pinned to exact versions for backend determinism; bumping them follows a benchmark-gated process documented in CLAUDE.md.

Useful checks:

./scripts/check_project_inputs.sh
./scripts/build_foundation_targets.sh macos
./scripts/build_foundation_targets.sh ios

More technical detail:

• CLAUDE.md — repo guide: build, architecture, engine invariants, dependency pinning, release policy, conventions
• docs/reference/cli.md — the headless vocello command-line tool
• docs/reference/privacy-storage.md — local storage and deletion details

Vocello ships a headless command-line tool, vocello, built from source alongside the app (it is not part of the app download). It drives the same local Swift + MLX engine in-process — no Python, no bundled weights — and serves two roles: scriptable local generation from the terminal, and the deterministic driver for the perf/quality benchmarks. It uses the models you install in the app (Settings → Model downloads); the CLI itself does not download weights.

./scripts/build.sh cli                 # build build/vocello
build/vocello <command> [options]      # run it (runs in place)

# Generate a clip (mode shortcut), or pipe a script in
build/vocello custom --variant speed --text "The train left at dawn."
echo "Hello there." | build/vocello generate --variant speed --stream --json

# Discover modes/speakers/models, then bulk synth (one model load)
build/vocello modes
build/vocello speakers list
build/vocello batch --file lines.txt --mode custom --variant speed --out-dir /tmp/batch

stdout is machine-readable (an output path, or JSON with --json); progress notes go to stderr. Full reference: docs/reference/cli.md.

Vocello is available under the MIT License.

Vocello builds on Qwen3-TTS, mlx-audio, MLX, and GRDB.swift.