Hugging Face released Phonon, a text-to-speech model that operates directly on personal devices without a network or server. The software reached more than 200,000 downloads in its first two weeks on the platform.
Phonon comes in sizes of approximately 65 megabytes and 75 megabytes. The models support English, French, German, Spanish and Portuguese through a collection of over 100 built-in voices.
Model Licenses and Benchmarks
Hugging Face structured the licensing differently depending on model size. The 31-million-parameter version uses the Apache 2.0 license for commercial use in English, while the 40-million-parameter version covers five languages under a non-commercial license.
On the Open TTS Leaderboard, Phonon ranked first in intelligibility for German and Spanish. The models placed in the top three across every evaluated language and topped the CPU streaming benchmark for time-to-first audio.
Inference and Framework
Developers use a dedicated inference framework named ptts to run the software across more than 10 backends. These include CPU architectures from Intel and Arm alongside various graphics processing units and neural processing units.
Transformers integration arrived from version 5.21 as a reference PyTorch implementation. The system generates audio latents autoregressively, feeding each output latent back as the subsequent input step during processing.
