Phonon is a small text-to-speech model from Gradium that runs entirely on-device. It requires no server or network, and the text being synthesized stays on the device. The weights are open: Phonon 31m is released under Apache 2.0 and can be used freely even in commercial products, while the other models, including Phonon 40m, are CC BY-NC and free for non-commercial work; commercial pricing for those is based on the number of users, not on generated speech volume. The release comes in two sizes, roughly 65 MB and 75 MB, covering English, French, German, Spanish and Portuguese, with more than 100 built-in voices and support for creating custom voices in Gradium Studio.

Getting a good voice to run on small hardware is the core problem Phonon addresses. Smaller models often sound poor, and high-quality models often do not fit on device. Phonon is based on the Pocket TTS lineage and uses flow matching to generate audio, which differs from the modeling techniques in other TTS systems; that approach is what enables the quality at small size, but it also made training numerically unstable at first, requiring several tricks to produce quality results across all five languages.

Phonon was announced in April, then built on for six months with partners and early users, which led to smaller models, more languages, and custom voice support through voice cloning and voice design. The launch included demos on an Apple Watch with Wi-Fi off, an iPhone with no signal, a robot named Lamp, an offline browser dino runner voiced by a character called Rex, a sci-fi game voiced locally, and Proust read aloud. Hugging Face built the Transformers integration, so Phonon is usable from Transformers from day one. Developers have added an ExecuTorch conversion for browser GPU use, a Node.js integration called Micdrop, and Spaces like smallest-s2s. Phonon was downloaded more than 200,000 times in its first two weeks.

On the Hugging Face Open TTS Leaderboard, which compares more than 9,000 TTS models across 23 languages using intelligibility (word error rate), speed (inverse real-time factor and time-to-first audio), and naturalness (UTMOS for English, KOMOS for Korean), Phonon ranks first in German and Spanish intelligibility and is in the top three in every language evaluated, next to models up to 100 times its size. Both fixed-voice and voice cloning are evaluated, with an additional speaker similarity metric for cloning. It tops the CPU streaming benchmark for time-to-first audio.

For inference, Gradium built a new framework called ptts. It targets the best performance across a wide range of devices and supports more than 10 backends: CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), and NPU (Apple Core ML, Snapdragon). Phonon is also available in Hugging Face Transformers from v5.21 as its reference PyTorch implementation. Where ptts runs a single stream on device, Transformers targets GPU throughput with batched generation and full-graph torch.compile, for example to synthesize speech datasets at scale. Phonon sits on top of Pocket TTS in Transformers as a modular model, with about 200 lines describing only what differs.

Phonon generates audio latents autoregressively: each step outputs a latent that is fed back as the next input, and the training loss is computed on those latents. The codec that maps audio to latents acts as an audio tokenizer, held by the processor next to the text tokenizer. A voice is not a recording but a tensor of speaker latents; voices come precomputed in gradium/phonon-voices, and the prompt layout (voice positions, text, then a token opening audio) ships with the checkpoint as a chat template. The code example shows loading the model, applying the chat template with an empty audio item where the voice goes, generating latents, and decoding them to audio.

Next steps include cloning a voice from a short recording with the speaker’s consent, or describing a new voice in plain text in Gradium Studio, then downloading the voice file and running fully on-device. Gradium also trains custom Phonon models with design partners: smaller models for tight memory budgets, additional languages such as Korean, Arabic and Japanese, and fine-tuning on product-specific vocabulary.