Gradium's Phonon TTS Runs Offline on Phones and Tops Open Leaderboard Intelligibility in German and Spanish

Gradium has published a deeper look at Phonon, its small text-to-speech model that runs on-device with no server and no network, roughly two weeks after opening the weights. The November 5, 2026 Hugging Face blog frames the release as voice that stays on the user’s hardware: typed text never leaves the device, and speech still works offline. Phonon was downloaded more than 200,000 times in its first two weeks and remains trending on Hugging Face.

Two sizes, five languages, and a split license

Phonon ships in two footprints of about 65 MB and 75 MB. Together they cover English, French, German, Spanish, and Portuguese, with more than 100 built-in voices. Custom voices come through Gradium Studio via cloning with speaker consent or plain-language voice design. Phonon 31m speaks English under Apache 2.0, so commercial products can use it freely. The other models—including Phonon 40m across all five languages—are CC BY-NC: free for non-commercial work, with commercial pricing based on user count rather than speech volume generated.

The stack sits in the Pocket TTS lineage and uses flow matching to generate audio, a modeling choice Gradium says is what lets quality fit in such a small model. Training was initially numerically unstable across the five languages and needed several tricks before quality held. The company also says Phonon handles phone numbers, dates, and codes more reliably than many TTS systems that stumble on those strings. Target hardware spans phones, laptops, single-board computers, and a browser tab.

Demos that only make sense offline

Launch demos lean into the no-network claim. An Apple Watch with Wi-Fi off walks through a coffee brew step by step. An iPhone with no signal counts workout reps. Autonomous’s Lamp robot wears a sorting hat and stays in character. An offline browser dino runner becomes a radio show hosted by Rex. A sci-fi game is voiced on the player’s machine, and Proust is read page by page. Hugging Face wired Transformers support from day one (v5.21); Pocket TTS landed first, and Phonon sits on top as about 200 lines describing only what differs.

Open TTS Leaderboard and the ptts backends

On Hugging Face’s Open TTS Leaderboard, Phonon ranks first on intelligibility in German and Spanish and in the top three in every language Gradium evaluated, next to models up to 100× its size. On CPU streaming, the models top time-to-first-audio. Gradium built a new inference framework, ptts, with more than ten backends: CPU on Intel, Arm, and WebAssembly; GPU on CUDA, Vulkan, Metal, and WebGPU; and NPU paths for Apple Core ML and Snapdragon. Transformers remains the path for GPU throughput—batched generation and full-graph torch.compile for dataset-scale synthesis—while ptts targets single-stream on-device runs.

Parameter counts published with the Transformers path put the Phonon 31m model at about 30.8 million parameters and the audio tokenizer at about 10.3 million. Voices are not raw recordings fed through the tokenizer at inference time; they ship as precomputed speaker latent tensors from the gradium/phonon-voices repository. The audio tokenizer is decoder-only in that path, and the prompt layout—voice frames, text, then an audio begin token—travels with the checkpoint as a chat template.

Roadmap: smaller customs, KR/AR/JA, vocab fine-tunes

Gradium says it already trains custom Phonon models with design partners: smaller footprints for tight memory budgets, more languages including Korean, Arabic, and Japanese, and vocabulary fine-tunes when a product needs its own terms. Independent third-party red-team or hallucination suites beyond the Open TTS Leaderboard metrics are not in the deep dive. What the post does nail down is a concrete open release: Apache English weights small enough for a watch, multilingual non-commercial siblings, a multi-backend local runtime, and leaderboard placements that put a ~31 million-parameter model beside systems two orders of magnitude larger.