Phonon
Gradium's open-weight, on-device text-to-speech models that run offline on phones, laptops, browsers and small boards.
At a Glance
Open-source on-device TTS runtime (MIT OR Apache-2.0) and Phonon 31m model (Apache 2.0). Phonon 40m models are CC BY-NC 4.0 and need a commercial arrangement for commercial use.
Engagement
Available On
Alternatives
Listed Nov 2026
About Phonon
Phonon is a family of open-weight text-to-speech models from Gradium, built to run on the device itself, including phones, laptops, a Raspberry Pi 5 and the browser. A Rust runtime, with no PyTorch required, ships alongside bindings for Python, Swift, Android, C, browsers and an OpenAI-compatible server. Once a model and voice are downloaded, no network connection is needed, so text and audio stay on the device.
What It Is
Phonon turns text into speech locally. It is aimed at offline assistants, accessibility tools and interactive experiences that need speech without a cloud round trip. Three models are available: Phonon 31m (English, 64 MB, Apache 2.0), Phonon 40m (English, French, German, Spanish and Portuguese, one model per language, 74.6 MB, CC BY-NC) and Phonon 40m multilingual (the same five languages in one model, CC BY-NC). Gradium says custom languages are built on request.
Where It Runs
The inference stack targets CPUs (x86-64 and ARM), Apple's Neural Engine through Core ML, Qualcomm Snapdragon NPUs through QNN, GPUs via CUDA, Metal, Vulkan or WebGPU, and browsers via WebAssembly. Streaming starts playback before a full utterance is ready, and speaking rate can be adjusted with the duration_delta condition. The README reports figures such as 22x real time on an iPhone 16 Pro Neural Engine and a 7 ms time to first byte for Phonon 40m on one MacBook Air core.
Voices and Integration
Each checkpoint includes a default voice, and a voice catalog on Hugging Face covers every supported language. Voices can be designed from a text prompt or cloned from a short recording in Gradium Studio or via scripts, then downloaded and used locally. Integration paths include the desktop ptts command, a Python package, an npm browser package, a Swift package, Android libraries, a Rust crate, a C library, a Docker image with an OpenAI-compatible API, and a Home Assistant app.
Licensing Notes
Code is MIT OR Apache-2.0, while model weights carry separate terms. Phonon 40m models are gated on Hugging Face, and commercial use of the CC BY-NC models requires contacting Gradium sales.
Community Discussions
Be the first to start a conversation about Phonon
Share your experience with Phonon, ask questions, or help others learn from your insights.
Pricing
Phonon Open Source (code and Phonon 31m)
Open-source on-device TTS runtime (MIT OR Apache-2.0) and Phonon 31m model (Apache 2.0). Phonon 40m models are CC BY-NC 4.0 and need a commercial arrangement for commercial use.
- Code licensed MIT OR Apache-2.0
- Phonon 31m (English) model under Apache 2.0
- Phonon 40m models under CC BY-NC 4.0 (gated; non-commercial)
- Runs on-device: phones, laptops, browsers, local services
- Rust runtime, no PyTorch needed
Commercial license (Phonon 40m)
Phonon 40m models are CC BY-NC 4.0; commercial use requires talking to sales.
- Licence CC BY-NC for Phonon 40m; commercial use via sales
Capabilities
Key Features
- Offline on-device text-to-speech
- Three models: Phonon 31m, 40m and 40m multilingual
- Five languages: English, French, German, Spanish, Portuguese
- Streaming audio with low time to first byte
- Adjustable speaking rate via duration_delta
- Voice design and voice cloning with downloadable voice files
- Runs on CPU, Core ML, Qualcomm QNN, CUDA, Vulkan, WebGPU and WebAssembly
- OpenAI-compatible speech server via Docker
- Bindings for Python, Rust, Swift, Android, C and browsers
- Desktop CLI for generating WAV files
