Phonon TTS: 65 MB On-Device Voice Model Tops German and Spanish Benchmarks
Gradium's Phonon TTS runs offline on phones and laptops at about 65 MB, ranks first in German and Spanish on Hugging Face's Open TTS Leaderboard, and is free under Apache 2.0.
Gradium’s Phonon, a text-to-speech model that runs entirely on your own device with no server and no network, ranks first in German and Spanish on Hugging Face’s Open TTS Leaderboard and lands in the top three in every language it was evaluated on, the company said in a blog post published November 5. The model was released two weeks earlier and has been downloaded more than 200,000 times since.
Phonon comes in two sizes, at about 65 MB and 75 MB. Together they speak English, French, German, Spanish and Portuguese, with more than 100 built-in voices. The company says the models also handle phone numbers, dates and codes, the details most TTS systems stumble on.
The pitch is privacy and offline use: what you type stays on the device. Gradium says Phonon runs on phones, laptops, single-board computers and in a browser tab.
Licensing splits by model
Phonon 31m speaks English and is Apache 2.0, so you can use it for free, including in commercial products. Phonon 40m covers all five languages under a CC BY-NC license, free for non-commercial projects. Gradium says commercial use of those models is priced by number of users rather than by how much speech you generate.
The models build on the Pocket TTS lineage and use flow matching to generate audio, an approach the company says differs from other TTS systems. Gradium notes training was initially numerically unstable and needed several tricks to produce quality results across all five languages.
First on time-to-first-audio on CPU
On the Open TTS Leaderboard, which evaluates models across 23 languages on intelligibility, speed and naturalness, Gradium says Phonon tops the CPU streaming benchmark for time-to-first audio. On intelligibility it ranks first in German and Spanish and in the top three in every language evaluated, next to models up to 100 times its size, according to the company.
Gradium built a new inference framework, ptts, for Phonon, supporting more than 10 backends across CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU) and NPU (Apple Core ML, Snapdragon). Phonon is also in Hugging Face Transformers from v5.21 as its reference PyTorch implementation, aimed at batched GPU generation rather than single-stream on-device use.
The launch shipped with demos, including Phonon talking through a coffee brew on an Apple Watch with Wi-Fi off and counting reps on an iPhone with no signal. Developers have also built an ExecuTorch conversion that runs Phonon on a visitor’s GPU in the browser and Micdrop for Node.js apps.
Voices can be cloned from a short recording with the speaker’s consent, or described in plain text in Gradium Studio and downloaded as a voice file that runs on-device. Gradium says it also trains custom Phonon models with design partners, including smaller models for tight memory budgets and new languages such as Korean, Arabic or Japanese.
The weights are on Hugging Face, and you can try Phonon in a browser tab from the company’s blog post.
Sources
- A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks. — Hugging Face Blog
Some links on this page may be affiliate links. If you buy through them we may earn a commission at no extra cost to you. See our affiliate disclosure.