Skip to content
Main Site News Console

Deep Dive into Phonon TTS: Introducing Benchmark-Topping, High-Quality On-Device Speech

· Hugging Face Translated
教程模型卡

Video 1

Two weeks ago, we open-sourced Phonon—a compact text-to-speech (TTS) model designed to run locally on your device. It requires no servers and no internet connection; your text never leaves your device.

We’ve been thrilled by the reception so far. Today, we want to share the story behind this release: why we built Phonon, what it can do, the accompanying demos, and how it performs on the Hugging Face Open TTS Leaderboard.

Why a voice should live on your device

Most voice AI runs in the cloud: every sentence makes a round trip between the client and a server. Phonon, by contrast, runs entirely on your device. Your text stays local, and the voice continues to work even when you’re completely offline. On top of that, we’ve open-sourced the model weights. Phonon 31M is released under the Apache 2.0 license, meaning you can freely use it in commercial products. The remaining models are under a CC BY-NC license, making them free for non-commercial projects; if you want to use them commercially, pricing is based on seat count rather than generated audio minutes.

Getting high-quality voice models to run smoothly on small devices has always been a tough challenge. Make the model small enough, and audio quality usually suffers; push for pristine audio quality, and the model becomes too large to fit on an edge device. We solved both problems at once.

We announced Phonon back in April and spent the next six months refining it alongside partners and early users. Their feedback helped shape this release: an even smaller footprint, support for more languages, and the ability to customize voices through Voice Cloning and Voice Design.

Phonon descends from the Pocket TTS family, utilizing Flow Matching to generate audio—a distinct modeling approach compared to most other TTS systems. While this technique enabled us to achieve remarkable audio fidelity in an exceptionally small model, it also introduced fresh training challenges. Early training suffered from numerical instabilities, and it took a combination of specialized techniques to reach high-quality generation across all five languages.

What we released

Phonon comes in two sizes, roughly 65 MB and 75 MB. Between them, they support English, French, German, Spanish, and Portuguese. The models include over 100 built-in voices, and you can also design custom voices in Gradium Studio. Furthermore, Phonon reliably handles critical edge cases where many other TTS systems stumble—such as phone numbers, dates, and code.

It runs directly on phones, laptops, single-board computers (SBCs), and even inside a browser tab—you can try it live right now. Phonon 31M supports English under the Apache 2.0 license, allowing developers to build freely on top of it (including for commercial applications). Phonon 40M covers all five languages under a non-commercial license.

Demos and first projects

Alongside this release, we showcased a series of demos created by community partners and our own team. On an Apple Watch with Wi-Fi disabled, Phonon guides you step-by-step through brewing coffee; on an iPhone with no cellular signal, it counts your workout reps and cheers you on; Autonomous’s robot Lamp wears a sorting hat and converses in distinct character voices; in the browser, the offline Chrome Dino game is reimagined as a radio show hosted by Rex; and there is even a sci-fi game with all voice acting generated entirely on the player’s device, alongside page-by-page readings of Proust.

The Hugging Face team completed the Transformers integration, enabling developers to use Phonon directly within the Transformers library from day one.

Developers have already started building with it. An ExecuTorch conversion runs Phonon directly in the browser using the visitor’s local GPU, [Micdrop](https

#教程#模型卡#AI语音#4Allapi.com

Published by the 4All API team

Original link:https://huggingface.co/blog/phonon-tts

Need a mainstream LLM API? 4All API gives you one key to call OpenAI, Anthropic, Google Gemini, Qwen, DeepSeek, and dozens more — at official-pass-through pricing, with enterprise-grade reliability, integrated in 5 minutes.

Sign up for the 4All API console →