An app that keeps speech synthesis entirely on-device on Apple Silicon Macs and iPhones. It runs Qwen3-TTS, Alibaba’s text-to-speech model, on Apple’s MLX framework with a native Swift runtime, so it needs no Python, no local server, and no cloud API. Write a script, pick a voice (or describe one in words), and you get an audio file. It is free and open-source under the MIT License, with no account signup and no usage-based billing.
The project was originally released as “QwenVoice,” but as of version 2.2.0 it was renamed to Vocello, and the repository moved from PowerBeef/QwenVoice to PowerBeef/Vocello. The official site has also moved to vocello.vercel.app; the old domain no longer resolves.
Key Features
- Fully offline operation: After the initial model download (via Hugging Face), speech is generated without any network connection. Scripts and generated audio stay in local app storage and are never uploaded unless you export them yourself
- Three ways to choose a voice: Nine built-in speaker presets (with eight delivery styles — Neutral, Calm, Whisper, Sad, Happy, Fearful, Angry, and Surprised), “Voice Design” for generating a voice from a plain-language description, and “Voice Cloning” for reproducing a voice from audio you have
- Voice cloning from recordings or files: You can either record directly with a microphone or import an existing audio file, and cloned voices can be saved and reused
- Project management for long scripts: Long scripts are automatically split into segments, and any segment can be regenerated from the history view. The design targets long-form work such as narration and audiobooks
- Automatic detection across ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian, with the input language detected automatically
- Native implementation without Python: Qwen3-TTS runs on a first-party Swift runtime built on mlx-audio-swift. On the Mac, the inference engine runs in a separate service process, so the model’s memory usage is less likely to affect the main app
- iPhone version in beta: The same local generation is available as a public TestFlight beta on iPhone 15 Pro or newer running iOS 26
Pricing
| Plan | Price | Key Features |
|---|---|---|
| Free (MIT License) | $0 | All features available. No subscription, no per-character billing, no account required |
Pricing is current as of August 2026. Check the official site for the latest information.
There are none of the “per-character” or “monthly character cap” limits typical of cloud speech synthesis services. The only costs are the Mac that runs the model and the bandwidth for the initial download.
Pros & Cons
✅ Pros
- With no usage-based billing, you can run long narration jobs and repeated retakes without worrying about cost
- Since audio never leaves the device, it suits confidential scripts and personal recordings
- The source code is published under the MIT License, so you can inspect it or build it yourself
- Even on an M2 Mac mini with 8 GB, generation runs at roughly 1.5–2.1× realtime, so waits are short
- Having presets, plain-language descriptions, and cloning all available gives you multiple routes to the voice you want
⚠️ Cons
- Limited to Apple Silicon Macs; it does not run on Intel Macs, Windows, or Linux
- Requires macOS 26 or later (the legacy QwenVoice 1.2.3 remains for macOS 15, but without the latest features)
- The initial model download requires a network connection
- The iPhone version is still a public TestFlight beta, and supported devices are limited to iPhone 15 Pro and newer
- As an individually developed open-source project, it offers no commercial support desk or SLA
- Voice cloning is meant to be used only with your own voice or a voice you have permission to use. Generated audio carries an inaudible watermark (AudioSeal), and publishing cloned-voice audio is expected to be disclosed as AI-generated
Comparison with Similar Services
| Criteria | Vocello | ElevenLabs | OpenAI TTS (API) | Kokoro TTS |
|---|---|---|---|---|
| Where it runs | On-device (local) | Cloud | Cloud | Local (self-managed) |
| Pricing model | Free, MIT License | Free tier + monthly subscription | Usage-based | Free, open-source |
| Voice cloning | Yes (mic recording or file import) | Yes (quality tied to higher plans) | No (preset voices only) | Limited |
| Ease of setup | Just download the Mac app | Runs in the browser | Requires an API key and code | Requires a Python environment |
| Requirements | Apple Silicon + macOS 26 or later | Web browser | Internet connection | GPU recommended |
The usual trade-off applies here as well: cloud services win on quality and convenience, local options win on cost and privacy. Vocello’s position is that it is a local option that still installs about as easily as any GUI app.
Who Is It For
- Anyone with an Apple Silicon Mac who wants to try speech synthesis for free first
- People who produce narration for videos or podcasts regularly and are concerned about usage-based costs adding up
- Corporate or personal use cases where scripts cannot be sent outside the organization
- Privacy-minded users who want to avoid uploading to cloud services
- People already running local LLMs and local generative AI who want speech synthesis on-device as part of that setup
Summary
Vocello (formerly QwenVoice) makes “speech synthesis that never leaves your machine” usable as an ordinary app by running Qwen3-TTS natively on Apple Silicon via MLX. The requirement of Apple Silicon plus macOS 26 or later is fairly strict, but if you meet it, the setup cost is a single download and there is no limit on characters or number of runs afterward. Just keep in mind that the name and repository have moved from QwenVoice to Vocello when searching for information.