Vocello is a voice studio for your Mac and iPhone. Write a script, choose a voice, and Vocello speaks it with Qwen3-TTS, an open text-to-speech model, running on your own Apple Silicon.
There is no account, no credits and no cloud. Your scripts, recordings, voices and takes stay on your device unless you export them.
- Built-in Voice. Nine voices in ten languages, with a delivery preset (calm, whisper, happy and more) or a few words of your own about how it should sound.
- Voice Design. Describe a voice, such as "a warm, clear narrator with a gentle British accent", hear it, and save the ones you like.
- Voice Cloning. Record or import a short clip of a voice you own or have permission to use. Voice Cloning follows the reference voice, so the clip sets its character, and it waits for your consent in Settings.
Vocello speaks Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian, and its interface comes in ten languages too, set apart from the speech language. Paste a whole chapter and it plans the script into segments that play one after another while the rest generates, then joins them into one file.
![]() |
![]() |
![]() |
![]() |
Once its model is loaded, Vocello makes speech about three times faster than you can listen to it, in every mode and at every script length, and a long script starts playing while the rest is still being made.
The real-time factor is the time it takes to generate one second of audio: at 0.30, a ten-second line takes three seconds. Each bar is the median of the warm takes in the latest benchmark records of the released build, on the benchmark Mac mini M6 (16 GB). The images in this section are regenerated from those records, so the numbers cannot drift from the evidence. → Benchmark history · How it is measured
On a Mac. Download Vocello, drag it to Applications and open it. Then open Settings > Models & Files and download the model of each mode you want to use.
On iPhone. Join the public beta on TestFlight. The beta still runs Vocello 3.0; the 3.1 update follows there.
In Terminal. Install the signed command-line package, which asks for an administrator password. From a new Terminal window, check it, download the Built-in Voice model (its id still says speed, a name kept from earlier versions) and play a first take:
vocello --version
vocello models install pro_custom_speed
vocello custom --text "The train left at dawn." --playvocello shares the Mac app's models, saved voices, settings and History, and answers in JSON for scripts. To let Claude Code speak through Vocello's MCP server:
claude mcp add --scope user vocello -- /usr/local/bin/vocello mcp| Platform | Requirements | Models |
|---|---|---|
| Mac | macOS 26.0 or newer, Apple Silicon, 8 GB of memory (16 GB or more recommended for long projects) | One model per mode |
| Command line | macOS 26.0 or newer, Apple Silicon | Shared with the Mac app |
| iPhone | iPhone 15 Pro or newer, iOS 26.0 or newer | One model per mode |
Models are never bundled. Each mode's model downloads once from Hugging Face, about 2.2 GB, and is verified before it is used. → What's new in 3.1.0 · CLI reference · All releases
Upgrading, and older Macs
History, saved voices and settings carry over from any earlier version. From 3.0, choose Vocello > Check for Updates; from 2.x, install the new DMG by hand once. Vocello then replaces the models you had with the new one for each mode, on its own on an unmetered network (about 4.7 GB for all three modes; Not Now stops it). The release notes have the details.
Vocello was called QwenVoice before 2.2.0, and old repository links redirect here. Macs on macOS 15 can use the legacy QwenVoice 1.2.3.
Note
Take quality is checked by machine, not by human listening. Automated checks catch gross failures (silence, a cut-off take, missing words, the wrong language), so a take can still carry a subtle defect, such as a swallowed syllable or a delivery that misses the requested style. Generate again for a different take. The release notes list the known limitations.
Vocello runs a 1.7-billion-parameter text-to-speech model natively on Apple Silicon: Swift 6 with strict concurrency, a first-party MLX runtime, and audio that streams with bounded memory. There is no Python, no server and no separate engine process.
One engine serves all three hosts. The talker turns your script into codec frames, the code predictor fills in each frame, and the speech decoder turns them into 24 kHz audio that streams, chunk by chunk, to the player and through a limiter, an atomic WAV writer, a watermark and an audio check. A take reaches History only when all of that has finished. → Architecture
An owned runtime, not a Python wrapper. VocelloQwen3Core is derived from mlx-audio-swift v0.1.2 and narrowed to what Vocello ships: about 36,000 of its 49,000 upstream lines removed, every remaining file tracked with its change, and a facade that keeps raw MLX types out of the apps. |
|
| One model per mode, tuned where it counts. Each mode runs a single mixed-precision model: 4-bit weights, with the talker's attention layers at 8 bits, where extra precision bought the most for the least weight. It was chosen by automated screens whose decision rules were written down before their takes were generated. → How the model was chosen | |
| Streaming that doesn't grow. Audio crosses an actor-owned, frame-bounded channel that never drops audio and suspends the producer instead of buffering, so the audio in flight is bounded by the channel, not by the length of the script. | |
| Launch-bound, not compute-bound. Fusing the code predictor's rotary embedding removed about 600 kernel launches per frame (26% faster warm), and compiling the per-frame predictor pass added another 8 to 11% with byte-identical output. → Optimization ledger | |
| Reproducible by construction. Every request carries its own seed and sampler state. Cancellation is typed and awaited, so a cancelled take never reaches History, and MLX errors arrive as typed failures instead of aborting the process. | |
| Proof before merge. Fixed-seed byte identity gates every scheduling change, published benchmarks are pass-only, privacy-checked records, and UI tests drive the real Mac app and a paired iPhone, never a Simulator. |
You need full Xcode on an Apple Silicon Mac with macOS 26 or newer; New Mac setup lists the pinned tools.
git clone https://github.com/PowerBeef/Vocello.git
cd Vocello
./scripts/regenerate_project.sh
./scripts/build.sh buildproject.yml generates the Xcode project, and the repository scripts are the build and test interface: scripts/dev.sh check runs the lint, contracts and tests your change touches. Start with CONTRIBUTING.md.
→ Architecture · CLI reference · Benchmarks · Testing runbook · Release notes · Roadmap
Vocello is available under the MIT License. It builds on Qwen3-TTS, MLX, mlx-audio-swift and GRDB.swift. Third-party models and assets keep their own terms; see the third-party attributions and the runtime's notices.





