Phonon 0.4.0
The first stable release of Phonon: streaming text-to-speech that runs on your device. The runtime is written in Rust and needs no PyTorch. It runs on desktops, servers, phones and in the browser.
The models
Phonon models are on Hugging Face. Without a model argument, every frontend downloads Phonon 31m on first use.
| Model | Repo | Languages | License |
|---|---|---|---|
| Phonon 31m | gradium/phonon-31m-en | English | Apache 2.0 |
| Phonon 40m | gradium/phonon-40m-en, -fr, -de, -es, -pt | one language each | CC BY-NC 4.0 |
| Phonon 40m multilingual | gradium/phonon-40m-multilingual | English, French, German, Spanish, Portuguese | CC BY-NC 4.0 |
The Phonon 40m models are gated. Accept the terms on the model's page (approval is automatic), then set HF_TOKEN or run hf auth login.
Voices come from the catalog gradium/phonon-voices, in all five languages. Name one with --voice and only that voice is downloaded. ptts --list-voices lists them.
Get it
- Command line:
uvx ptts --lang en "Hello from Phonon." -o speech.wav,brew install gradium-ai/tap/ptts, or the archives below. - Python:
pip install ptts - Browser:
npm install phonon-tts - Rust:
cargo add ptts --features hf,hub - Swift: add
https://github.com/gradium-ai/xn-ptts.gitin Xcode. - Server:
ghcr.io/gradium-ai/ptts-openai-server:0.4.0, an OpenAI-compatible speech API that also speaks the Wyoming protocol for Home Assistant. - Android: a C library with a Kotlin wrapper on the CPU, and a QNN library for Snapdragon NPUs, both built from source.
Model bundles in this release
Ready-made bundles of Phonon 31m, the default model, with the English voices of the catalog:
phonon-31m-en-coreml.zip: for the Swift package (Core ML, iPhone and Mac).phonon-31m-en-qnn.zip: for the Android QNN library, with context binaries for the SM8650, SM8750, SM8850 and QCS8275.
SHA-256:
10f3a99fd39256db44f826267ef21d78cca959b16c0554926344c09fde27b436 phonon-31m-en-coreml.zip
9da103025d88682428e350f7d65254bfa672d671a9a7e643fb4527e370007722 phonon-31m-en-qnn.zip
For a Phonon 40m model, accept its terms on Hugging Face, then export a Core ML bundle with export_coreml or build a QNN bundle with the tools in qnn/export.