Most AI voices live in the cloud: every sentence you want spoken goes to a server and comes back as audio. Phonon, a family of text-to-speech (TTS) models from Gradium hosted on Hugging Face, works the other way. It is small enough to run on a phone, a laptop CPU, or inside a browser tab, with no internet connection. A Hugging Face blog post on November 5, 2026 gave a full look at the models.

Coffee summary

  • FACT: The English model, phonon-31m-en, has 31 million parameters, a 64.5 MB weights file, and an Apache 2.0 license that allows commercial use.
  • FACT: Larger 40M models cover English, French, German, Spanish, and Portuguese, but they are licensed CC BY-NC 4.0 (non-commercial).
  • FACT: Output is 24 kHz mono speech that starts streaming while it is still being generated.
  • CLAIM: Gradium says Phonon tops on-device TTS benchmarks. We have not seen independent tests yet.

What happened

Gradium first opened the Phonon models about two weeks ago and has now published a detailed launch post. The models build on Kyutai’s Pocket TTS and Mimi audio codec. They come with more than 100 built-in voices and handle tricky text like phone numbers, dates, and codes. Gradium also built a runtime called ptts that supports more than 10 hardware backends, from regular CPUs and WebAssembly to Apple and Snapdragon chips. Phonon can also run as a local server with an OpenAI-compatible API.

Why it matters

On-device speech has three big advantages. It is private, because your text never leaves the device. It works offline. And it responds instantly, because there is no network round trip. That makes it a good fit for screen readers, offline assistants, games, and apps that read content aloud.

Who should care

  • App and game developers who need a voice without paying per character to a cloud service.
  • Accessibility builders working on screen readers and reading tools.
  • Home Assistant users, since Phonon can act as a local voice.

Limitations

Only the English 31M model is free for commercial use; the multilingual models need a paid license for commercial products. Each model speaks only its listed languages, and text in other languages comes out with the wrong pronunciation. Long texts are split into sentences, so intonation does not carry across a whole paragraph. Voice cloning goes through Gradium Studio, and you should only clone a voice with the speaker’s consent.

What to do next

  1. Try it from a terminal with Python’s uv installed: uvx ptts --lang en --quant q8 --voice Easton "Hello from Phonon." -o speech.wav
  2. Or in Python: pip install ptts, then load hf://gradium/phonon-31m-en.
  3. Set the lang option to match your text, so numbers and dates are read correctly.
  4. Check the license before shipping: Apache 2.0 for phonon-31m-en, non-commercial for the 40M models.

AIImpish Take

A natural-sounding voice in about 65 MB that runs offline is a big deal for privacy and cost. If you build in English, phonon-31m-en is worth testing today. For other languages, read the license first.

Related on AIImpish

Sources: Hugging Face blog, A Closer Look at Phonon TTS; Hugging Face, gradium/phonon-31m-en model card.