Gradium's Phonon puts benchmark-level text-to-speech on phones and laptops

Gradium, a speech company, has released Phonon, a small text-to-speech model that runs entirely on the user's device. A launch post on the Hugging Face blog, published on 5 November 2026 about two weeks after the models were opened up, describes how it was built and how it scores. The text is never sent to a server, and the voice works without a network connection. Hugging Face built the integration into its Transformers library.

The main reason it matters is size. The English model has about 31 million parameters, and its quantized file is 64.5 MB. On Hugging Face's Open TTS Leaderboard it ranks first for intelligibility in German and Spanish and in the top three in every language tested. According to the post, it does this alongside models up to 100 times larger.

What was released

Phonon comes in two versions. Phonon 31m covers English only and is licensed under Apache 2.0, which allows commercial use. Phonon 40m covers English, French, German, Spanish and Portuguese under a non-commercial license. Gradium says commercial use of the non-commercial models is priced by number of users rather than by the amount of speech produced.

The models come with more than 100 built-in voices. Gradium also offers voice cloning from a short recording, which requires the speaker's consent, and voice design from a text description, both through its Gradium Studio product. The voices are stored as precomputed speaker tensors, not as raw recordings.

According to the model card, Phonon builds on Pocket TTS and the Mimi audio codec from Kyutai. A flow-matching transformer predicts audio latents in 80 millisecond frames, and the Mimi decoder turns them into 24 kHz mono audio. The audio streams while the model is still generating. The blog post says training was numerically unstable at first and needed several techniques before quality held up in all five languages.

How it scores

The Open TTS Leaderboard covers 23 languages. It measures intelligibility by word error rate, speed by real-time factor and time to first audio, and naturalness with automated scores. For latency, Phonon tops the CPU streaming benchmark. At the 95th percentile, Phonon 31m English produces its first audio in 28 milliseconds and Phonon 40m English in 38 milliseconds. Pocket TTS, the model it descends from, takes 165 milliseconds.

The model card gives English results against other small open models. Phonon 31m has a word error rate of 1.5 percent on a hard test set, compared with 3.5 percent for Kokoro, 8.7 percent for NeuTTS and 28 percent for Magpie. The larger Phonon 40m scores 2.3 percent on that set. On the easier Seed-TTS set the models are close together: Phonon 31m scores 0.51 percent and Kokoro and Magpie 0.45 and 0.46 percent. In the card's Elo preference rating, Phonon 40m leads with 2117, followed by Phonon 31m at 2098 and Kokoro at 2043.

These figures come from Gradium's own post and model card and from a leaderboard hosted by Hugging Face, which helped with the release.

Where it runs

Gradium released a new inference framework, ptts, with more than ten backends. These cover CPUs (Intel, Arm, WebAssembly), GPUs (CUDA, Vulkan, Metal, WebGPU) and neural processors (Apple Core ML, Snapdragon). The post says the model runs on phones, laptops, single-board computers and in a browser tab. The model card adds that it can run as a local server compatible with OpenAI's API and as a voice for Home Assistant. From Transformers version 5.21, Phonon is included as a reference PyTorch implementation meant for batched GPU work such as generating speech datasets.

The demos in the post include an Apple Watch guiding someone through brewing coffee offline, an iPhone counting exercise reps with no signal, a robot that speaks in character, a game voiced on the player's own machine, and the reading of Proust page by page. According to Gradium, the models were downloaded more than 200,000 times in the first two weeks.

The model card lists limits. Each model handles only its listed languages, and other languages come out with the wrong pronunciation. Long texts are split into sentences, so intonation does not carry across a whole passage. The Transformers tokenizer does not spell out numbers, amounts, symbols or addresses. The card also prohibits impersonating real people, cloning voices without consent and misleading listeners about who is speaking.

What it means for work

Speech that used to need a cloud service can now run on the device that plays it. That covers screen readers, voice prompts in apps and devices, game dialogue, and reading articles or books aloud. For developers, this removes per-request API costs and network delay, and user text stays on the device. That matters for private assistants and for regulated settings. For voice actors and narrators, the pressure comes mostly in volume work: the model ships with 100 ready-made voices and can clone a voice from a short recording. Long-form narration that depends on intonation across a whole passage is less exposed for now, because the model reads sentence by sentence. Still unclear are how well the high benchmark scores hold up in real products, and how consent rules for cloned voices will be enforced once the software runs offline on someone else's device.