Gradium’s Phonon TTS Runs Entirely On-Device, No Cloud Needed

Gradium has published a detailed look at Phonon, its family of on-device text-to-speech models, two weeks after the public launch. The headline numbers: more than 200,000 downloads in the first two weeks, first-place intelligibility rankings in German and Spanish on Hugging Face’s new Open TTS Leaderboard, and the fastest time-to-first-audio on CPU streaming benchmarks — all from models that weigh in at roughly 65 to 75 megabytes and need no internet connection to run.

Phonon comes in two versions. Phonon 31m covers English and is licensed under Apache 2.0, meaning it is free to use in commercial products without royalties or usage fees. Phonon 40m extends support to French, German, Spanish and Portuguese but carries a CC BY-NC license — free for non-commercial work, but developers building a paying product will need to negotiate terms with Gradium directly. Pricing, the company says, is based on user count rather than the volume of speech generated. Both models ship with more than 100 built-in voices, and the Gradium Studio tool supports voice cloning from a short recording and voice design from a plain-text description.

How It Works Under the Hood

Phonon is built on the Pocket TTS lineage and uses a technique called flow matching to generate audio — a departure from the autoregressive or diffusion-based approaches used by most other TTS systems. Gradium says this architecture is what made it possible to reach high audio quality at such a small model size, though it came with trade-offs: training was initially numerically unstable and required specific mitigation techniques before the models produced consistent results across all five languages.

For deployment, Gradium built a custom inference framework called ptts that supports more than 10 hardware backends, including CPU targets (Intel, Arm, WebAssembly), GPUs (CUDA, Vulkan, Metal, WebGPU) and NPUs (Apple Core ML, Snapdragon). That breadth is deliberate — the goal is a single model that runs on a phone, a Raspberry Pi, a laptop, or a browser tab without code changes. Phonon is also integrated into Hugging Face Transformers starting with version 5.21, giving developers access to batched generation and full-graph torch.compile for large-scale use cases like synthesizing speech datasets.

Where It Stands Against the Competition

The on-device open-source TTS field already had two notable incumbents before Phonon arrived: Kokoro-82M and Piper. Kokoro reached the top of the Hugging Face TTS Arena leaderboard and is Apache 2.0 licensed, but it synthesizes a full sentence before emitting any audio, which means its first sound arrives roughly a second later than Phonon’s. Piper supports desktop and Raspberry Pi but lacks official iOS and browser runtimes. Phonon explicitly targets all of those platforms through ptts.

The licensing picture is worth reading carefully. Kokoro and Pocket TTS are English-only. Piper’s original MIT-licensed repository was archived in October 2025, with active development shifting to a GPL-3.0 fork — still usable commercially, but copyleft terms matter for closed-source products. Phonon’s Apache 2.0 English model is one of the cleaner commercial licenses in the space right now. The catch, flagged by community commenters on Hugging Face, is that the multilingual capability — which is where Phonon’s benchmark advantages are most pronounced — sits behind CC BY-NC. One commenter put it plainly: the family framing can obscure the fact that most of Phonon’s languages are not available for commercial work without a separate agreement.

Why This Matters If You’re Building Something

For students, recent graduates and self-taught developers, the practical significance of Phonon is straightforward: it removes the two biggest obstacles to adding speech to a project. Cloud TTS APIs from providers like ElevenLabs, OpenAI, or Google charge per request, which scales badly for consumer apps with real user numbers. Heavier open-source alternatives typically demand GPU infrastructure that is expensive to run continuously. Phonon 31m runs on a laptop CPU, starts producing audio in milliseconds, and costs nothing to deploy at scale under its Apache 2.0 license.

That combination unlocks project types that were previously impractical at the student or indie level. A mobile game where non-player characters speak dynamically, an offline accessibility tool, a voice-enabled study app that works without Wi-Fi — all of these require a TTS model that runs locally, generates quickly and carries no per-request cost. Developers who already use the Transformers library can plug Phonon in without learning a new API. The browser-compatible runtime means voice features can ship as fully client-side web apps with no backend and no API key to manage.

The main limitation to keep in mind: if you need French, German, Spanish or Portuguese and plan to charge for your product, the 40m model’s CC BY-NC terms mean you will need to contact Gradium before shipping. For English-only applications, the 31m model has no such restriction.

Source: Hugging Face