Vocello runs Qwen3-TTS on Apple Silicon. Custom voices, voice design, and voice cloning, all on-device. No accounts. No credits.
Apple Silicon. macOS 15+.
Three modes
Pick a preset speaker. Type a script. Hear it back, streaming, before you've stopped scrolling.
Four English speakers ship with the app: Ryan, Aiden, Serena, Vivian.
Ryan
Steady, mid-range.
“Hey there. Welcome back to the studio.”
Aiden
Lower, warmer.
“Long-form, considered. Steady tempo, low register.”
Serena
Bright, articulate.
“Crisp, precise, alert. Like a bell at dawn.”
Vivian
Calm, mature.
“A patient narrator. Considered, without urgency.”
Audition, refine, save. Voice Design is a destination of its own, not a parameter panel buried inside another flow.
Saved voices stay in a local library. Open one in Voice Cloning whenever you want to start from it.
“A weathered storyteller. Slow tempo, low register.”
“Energetic, fast-paced. Calls plays in real time.”
“Intimate and close. Soft consonants, slow breath.”
“Confident, warm, articulate. Pauses on key terms.”
WAV, MP3, AIFF, M4A, FLAC, or OGG. Add a transcript when accuracy matters. Generation runs locally; the reference never leaves the machine.
A short clip of the voice you want.
The model speaks your text in that voice. Locally, on your Mac.
Reference formats. WAV · MP3 · AIFF · M4A · FLAC · OGG. Optional transcript improves accuracy.
Only clone voices you own or have permission to use.
Install
Pull Vocello-macos26.dmg from the latest GitHub release.
Drag Vocello.app to Applications. Open it.
Pick Custom Voice, Voice Design, or Voice Cloning from inside the app. Models stay local.
Type, choose a voice, hit generate. Live preview plays as the audio renders.
Under the hood
A native SwiftUI app shell hosts an MLX-backed engine in an isolated XPC helper. Each generation runs out-of-process, so a heavy job never freezes the UI and a model crash never takes the app with it. The iPhone target uses an engine extension for the same separation on iOS.
Process boundary
SwiftUI app shell
Native macOS UI. Holds the script, the player, the saved voices.
XPC helper
Out-of-process boundary. Crashes here never reach the UI.
MLX engine
Inference on Apple Silicon. Local model weights, GPU + Neural Engine.
Models
8-bit for quality, 4-bit for speed.
Privacy
Inference runs locally. Models, prompts, generated audio, and every saved voice sit in your home folder, where Finder can see them.
Generation history is a local SQLite file. There is no account, no credit meter, no telemetry that leaves the machine, no external service handling your text or your audio.
You can delete a voice the same way you delete any other file: open the folder, drag it to the Trash.
~/Library/Application Support/Vocello/├── models/ // downloaded model weights├── outputs/│ ├── CustomVoice/│ ├── VoiceDesign/│ └── Clones/├── voices/ // your saved voices└── history.sqlite // every generation, locally
macOS 15.0+, Apple Silicon, 8 GB+ RAM.
Signed, notarized, and stapled by the official build pipeline.
Local debug or pre-release artifacts may still carry quarantine. Clear it from Terminal:
xattr -cr "/Applications/Vocello.app"