Gradium’s Phonon text-to-speech models generate speech on-device, without a server or network connection, and are available in two small sizes across five languages. The release pairs that privacy-friendly setup with open weights, broad hardware support and strong results on Hugging Face’s Open TTS Leaderboard, according to the project’s launch account.
Hugging Face Watch analysis
What happened
Phonon comes in models of about 65 MB and 75 MB, covering English, French, German, Spanish and Portuguese. Gradium says they include more than 100 built-in voices and can also use custom voices created through voice cloning or voice design.
The licensing differs by model: Phonon 31m, which speaks English, is Apache 2.0 and can be used commercially. The five-language Phonon 40m is under a non-commercial licence. Gradium says its new inference framework, ptts, supports more than ten backends across CPU, GPU and NPU hardware. Phonon is also available in Transformers, where the implementation is aimed at GPU throughput and batched generation.
Our top picks
- Speech stays on the device
Generation needs no server or network connection, keeping the typed text on the user’s device. - Small models, five languages
The two models are about 65 MB and 75 MB and together support English, French, German, Spanish and Portuguese. - Licensing is not one-size-fits-all
The English-only 31m model is Apache 2.0; the five-language 40m model is for non-commercial use. - A wide hardware spread
Gradium says ptts supports CPU, GPU and NPU backends, including Arm, WebAssembly, CUDA, Metal and Apple Core ML. - Promising leaderboard results
Gradium says Phonon ranked first in German and Spanish intelligibility and in the top three for every language it evaluated. - A quick start for developers
The project says Transformers support is available, alongside ptts for on-device inference.
Why it matters
Small speech models that can run locally could make voice features more practical on phones, laptops and other devices, including when connectivity is poor or privacy matters. The trade-off is that “open” does not mean every version has the same commercial terms, so builders will need to check the licence before choosing a model.
The leaderboard results are Gradium’s account of its evaluations, not a guarantee that every device or voice will sound equally good. Still, putting compact models alongside a dedicated inference framework and multiple hardware backends gives developers something more useful than a promising demo: a route to trying local speech in their own products.
Our read
This is a substantial push towards speech generation that does not need to phone home. The model sizes and hardware support make the practical proposition easy to grasp; the split licences are the detail to read before building a business around it. If local speech is on your roadmap, start with the model that fits your language and licence needs, then test it on the hardware you actually plan to ship.
What to watch
- How Phonon sounds and performs across different devices and real-world workloads.
- Whether the leaderboard results hold up in broader comparisons and across more languages.
- Which integrations and custom-voice options Gradium develops next.
Discussion spark: Would you choose a smaller on-device voice model for privacy and offline use, or stick with cloud speech if it sounds better?
Sources and evidence
- A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks. (5 November 2026, 00:00 UTC)
- From Chatbot to Phone Agent: Adding Voice to Your LLM Support Bot (5 November 2026, 00:00 UTC)
not affiliated with or endorsed by Hugging Face