Hugging Face Launches Phonon TTS for On-Device Speech

Hugging Face Phonon TTS launch on-device speech

Hugging Face just dropped Phonon, a family of tiny text-to-speech models that run completely on-device-no server round-trip required. The Gradium team pushed the release on November 5, 2026, wrapping up six months of polish with partners and early testers after first teasing the project back in April. You get two flavors: Phonon 31m handles English under an Apache 2.0 license (free for commercial work), while Phonon 40m adds French, German, Spanish, and Portuguese under a Creative Commons non-commercial license. Both weigh in around 65 to 75 megabytes and ship with over a hundred built-in voices; if you need something bespoke, Gradium Studio lets you cook up custom ones.

This release tackles a classic headache in speech synthesis. Small models usually sound robotic, and the good-sounding ones are too hefty for phones or embedded boards. Phonon dodges that trade-off by leaning on the Pocket TTS lineage, which uses flow matching instead of the usual autoregressive or diffusion tricks. That architectural bet let them cram serious quality into a minuscule footprint, though it brought numerical instability during training that took a bag of engineering fixes to smooth out across all five languages.

Phonon TTS privacy local processing no server

Because everything runs locally, your text never leaves the device. That’s a big deal for privacy-sensitive apps, spotty connectivity, or anyone trying to dodge per-character API fees. Phonon runs on phones, laptops, single-board computers, and even inside a browser tab via WebAssembly or WebGPU. The launch demos show the range: an Apple Watch guiding a coffee brew with Wi-Fi off, an iPhone counting reps without a signal, a robot named Lamp wearing a sorting hat and staying in character, a browser-based dino runner narrated like an old radio drama, a sci-fi game voiced entirely on the player’s machine, and a page-by-page reading of Proust.

Developers have two integration paths. Transformers added support in version 5.21, giving you a native PyTorch implementation geared for batched generation and large-scale dataset synthesis. Separately, a framework called ptts targets single-stream, low-latency performance across more than ten backends-Intel and Arm CPUs, CUDA, Vulkan, Metal, WebGPU, Apple Core ML, and Qualcomm Snapdragon NPUs. Inside Transformers, Phonon sits atop Pocket TTS as a modular add-on of roughly two hundred lines. The model generates audio latents autoregressively, treating latents like a language model treats text tokens. An audio tokenizer (bundled in the processor with the text tokenizer) maps between waveforms and latents. Voices live as precomputed speaker latent tensors rather than raw recordings, and the prompt layout follows the chat-template convention used by other multimodal models in Transformers.

Phonon TTS adoption benchmarks Hugging Face leaderboard

Early traction has been wild. The models cleared two hundred thousand downloads in their first two weeks and are still trending on the Hugging Face Hub. Community contributions are already piling up: an ExecuTorch conversion running Phonon on a visitor’s GPU in the browser, a Node.js wrapper dubbed Micdrop, and a handful of Spaces built on top.

Benchmark numbers on the Hugging Face Open TTS Leaderboard back up the hype. The leaderboard-a fresh project evaluating over nine thousand TTS models across twenty-three languages-measures intelligibility via word error rate, speed via inverse real-time factor and time-to-first-audio, and naturalness via UTMOS for English and KOMOS for Korean. Phonon grabs first place in German and Spanish intelligibility and lands in the top three for every language tested, going toe-to-toe with models up to a hundred times larger. On the CPU streaming benchmark, it leads on time-to-first-audio, the metric that actually matters for responsive voice interfaces.

Phonon TTS commercial licensing English multilingual models

For builders, the takeaway is simple: high-quality, multilingual, on-device speech is finally practical-no licensing fees for English apps and a clear commercial path for the bigger model. The Transformers integration means existing PyTorch workflows can adopt Phonon with barely any friction, while ptts gives you a lightweight runtime for edge deployment. If you’re weighing TTS options, check the Open TTS Leaderboard’s Listen tab to audition Phonon side-by-side with larger cloud models. That combo of small size, strong benchmarks, flexible licensing, and broad hardware support makes Phonon a credible default for any product that needs to speak without phoning home.

Frequently asked questions

What is Phonon and when was it released?

Phonon is a family of tiny text-to-speech models from Hugging Face that run entirely on-device. The Gradium team released it on November 5, 2026, after six months of development with partners and early testers.

What languages and licenses do the two Phonon models support?

Phonon 31m handles English under an Apache 2.0 license (free for commercial use). Phonon 40m adds French, German, Spanish, and Portuguese under a Creative Commons non-commercial license. Both are 65-75 MB with 100+ built-in voices.

How does Phonon achieve high quality in such a small model size?

Phonon uses the Pocket TTS lineage with flow matching instead of autoregressive or diffusion approaches. This architecture packs strong quality into a tiny footprint, though it required engineering fixes to resolve numerical instability across all five languages.

Where can Phonon run and what are the privacy benefits?

Phonon runs on phones, laptops, single-board computers, and in browsers via WebAssembly or WebGPU. Because everything runs locally, your text never leaves the device-ideal for privacy-sensitive apps, offline use, or avoiding per-character API fees.

What integration options exist for developers?

Two paths: Transformers 5.21+ provides a native PyTorch implementation for batched generation and dataset synthesis. The ptts framework targets single-stream, low-latency performance across 10+ backends including CPUs, GPUs, NPUs, and WebGPU.

Source: Hugging Face