Hugging Face releases Phonon TTS, an on-device text-to-speech model with Apache 2.0 weights
Phonon TTS runs entirely on-device, supports five languages, and ranks first in intelligibility for German and Spanish on the Hugging Face Open TTS Leaderboard.
1 source · cross-referenced
- Hugging Face released Phonon TTS, an on-device text-to-speech model with Apache 2.0 weights for English and CC BY-NC for four other languages.
- Phonon comes in two sizes (~65 MB and ~75 MB), supports English, French, German, Spanish, and Portuguese, and includes over 100 built-in voices.
- On the Hugging Face Open TTS Leaderboard, Phonon ranks first in German and Spanish intelligibility and tops CPU streaming benchmarks for time-to-first audio.
- Phonon uses flow matching and runs on phones, laptops, single-board computers, and in browsers via a new inference framework, ptts, with support for 10+ backends.
Hugging Face released Phonon TTS, a small text-to-speech model designed to run entirely on-device without requiring server-side processing or network connectivity. The model’s weights are available under Apache 2.0 for the English-only variant (Phonon 31m) and a non-commercial license (CC BY-NC) for the multilingual variant (Phonon 40m), which supports English, French, German, Spanish, and Portuguese.
Phonon is offered in two sizes, approximately 65 MB and 75 MB, and includes more than 100 built-in voices. Users can also create custom voices using Gradium Studio. The model addresses challenges in balancing model size and audio quality by leveraging flow matching, a technique distinct from modeling approaches used in other TTS systems, which enabled high-quality speech generation within a compact footprint.
On the Hugging Face Open TTS Leaderboard, Phonon ranks first in intelligibility for German and Spanish and is in the top three across all 23 evaluated languages. It also leads CPU streaming benchmarks for time-to-first audio, outperforming models up to 100 times its size in speed metrics.
Phonon runs on a variety of hardware, including phones, laptops, single-board computers, and browsers. Hugging Face developed a new inference framework, ptts, to optimize performance across more than 10 backends such as CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), and NPU (Apple Core ML, Snapdragon). The model is also integrated into the Transformers library starting from version 5.21 as a reference PyTorch implementation.
Since its launch, Phonon has been downloaded over 200,000 times and has been adopted in projects like ExecuTorch for browser-based GPU inference, Micdrop for Node.js applications, and Hugging Face Spaces such as smallest-s2s. Demos accompanying the release include an Apple Watch guiding coffee brewing without Wi-Fi, an iPhone counting reps offline, and a sci-fi game voiced locally on the player’s device.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79