Gradium Phonon TTS Tops Hugging Face Benchmarks With Offline Speech
Gradium's Phonon brings private, offline speech to small devices, tops key Hugging Face tests, and passes 200,000 downloads in its first two launch weeks.
Summary
By November 5, 2026, Gradium's open-weight Phonon TTS had been public for two weeks after an April announcement and six months of work with partners and early users. It runs locally without servers or networking on phones, laptops, single-board computers and browsers, keeping text on device. Two models of about 65 MB and 75 MB offer more than 100 voices, custom voice cloning and design, and handling for phone numbers, dates and codes. Phonon 31m is English-only under Apache 2.0 for free commercial use; Phonon 40m covers English, French, German, Spanish and Portuguese under CC BY-NC, with commercial pricing tied to users rather than generated speech.
Among more than 9,000 models on Hugging Face's new Open TTS Leaderboard, which tests 23 languages, Phonon ranks first for German and Spanish intelligibility, top three in every evaluated language against models up to 100 times larger, and tops CPU streaming time to first audio. It exceeded 200,000 downloads in two weeks and trended on Hugging Face. Offline demos ran on Apple Watch and iPhone, plus Autonomous's Lamp robot, Rex browser radio, a local science-fiction game and a Proust reader; Hugging Face delivered day-one Transformers support, while ExecuTorch, Micdrop and smallest-s2s projects followed.
Phonon uses Pocket TTS flow matching and autoregressive audio latents; training was initially numerically unstable and required fixes across all five languages. The ptts inference framework supports more than 10 backends spanning Intel, Arm and WebAssembly CPUs; CUDA, Vulkan, Metal and WebGPU GPUs; and Apple Core ML and Snapdragon NPUs. Transformers v5.21 provides the PyTorch reference for batched GPU generation; Phonon 31m's model has 30,788,354 parameters and its audio tokenizer has 10,305,313. Gradium Studio can create consent-based cloned or described voices for local use, while design partners can request smaller models, Korean, Arabic or Japanese training, and vocabulary fine-tuning.
Positives
- More than 200,000 downloads arrived during Phonon's first two weeks, with the models continuing to trend on Hugging Face.
- First-place German and Spanish intelligibility scores put Phonon ahead of some models up to 100 times larger.
- Phonon tops the Open TTS Leaderboard's CPU streaming benchmark for time to first audio.
- Apache 2.0 licensing lets developers use the English Phonon 31m model free in commercial products.
- More than 10 ptts backends extend local inference across Intel, Arm, WebAssembly, CUDA, Vulkan, Metal, WebGPU, Core ML and Snapdragon hardware.
Risks & concerns
- Phonon 40m uses a CC BY-NC license, requiring separate user-based pricing for commercial products.
- Five languages are currently supported, while Korean, Arabic and Japanese require custom training with Gradium.
- Phonon 31m supports only English despite offering the release's most permissive commercial license.
- Flow-matching training was initially numerically unstable and needed several fixes to deliver quality across all five languages.
- Hugging Face's Open TTS Leaderboard is still a very new benchmark despite evaluating more than 9,000 models across 23 languages.