Hugging Face and Gradium Recap Phonon Two Weeks After Its Release: Small On-Device TTS Models Reported by the Coauthors at More Than 200,000 Downloads and Top Intelligibility in German and Spanish on the Open TTS Leaderboard, but Only the 31m English Model Is Apache 2.0 and the rankings are the coauthors’ leaderboard reading and performance figures are Gradium measurements

On November 5, 2026, the Hugging Face blog published A Closer Look at Phonon, written by people from Gradium and Hugging Face. It looks back two weeks at Phonon, a family of small text-to-speech (TTS) models that Gradium open-sourced on October 20, 2026, and that run on the user's own device with no server and no network. This is a launch retrospective, not a new release, and a partner article, not an independent review. The Gradium release post supplies the model table and benchmark detail used below.

The models come in two sizes. Phonon 31m speaks English, is about 64 MB in its 8-bit form and uses the Apache 2.0 license, so it can be used in commercial products. Phonon 40m is about 74.6 MB, covers English, French, German, Spanish and Portuguese either as one model per language or as a single multilingual model, and uses the CC BY-NC license, which allows non-commercial use only. The post says commercial use of those models is priced by the number of users, not by how much speech is generated. There are more than 100 built-in voices, and users can clone a voice from a short recording, with the speaker's consent, or design one from a text description in Gradium Studio. The models handle numbers, dates, phone numbers and codes through text normalization. Phonon belongs to the Pocket TTS lineage from Kyutai and generates audio with flow matching.

The post reports adoption and ranking. It says Phonon was downloaded more than 200,000 times in its first two weeks and is still trending on Hugging Face. On the Hugging Face Open TTS Leaderboard, which compares open models across 23 languages on intelligibility (word error rate), speed and naturalness, the authors say Phonon ranks first in German and Spanish on intelligibility and in the top three in every language they evaluated, next to models up to 100 times its size, and that it tops the CPU streaming benchmark for time to first audio. These readings of the leaderboard are the authors' own, and we did not check the live leaderboard.

Gradium's launch post adds its own measurements. On a test set of 500 hard sentences per language, collected from cases that developers said their current models get wrong, English word error rate was 1.70% for Phonon 31m, 2.22% for Phonon 40m and 2.11% for the multilingual model, against 3.50% for Kokoro, 7.40% for Kitten TTS, 8.71% for NeuTTS Nano and 27.51% for Magpie. The transcripts were made by a multimodal model and the error rate was scored by a language model, and the test set was built from the failures of other systems. On speed, Gradium reports that on a MacBook Air M5 CPU Phonon starts speaking in about 7 milliseconds against 517 for Kokoro, which does not stream, so its first audio is the whole sentence. On a Raspberry Pi 5, Phonon 31m starts in 38 milliseconds and runs 3.7 times faster than real time. Gradium also ran blind listening tests with 1,189 raters and 17,270 votes against Kokoro, Magpie, NeuTTS and Pocket TTS.

Phonon ships with its own inference framework, ptts, which supports more than 10 backends across CPU, GPU and NPU, including WebAssembly, CUDA, Vulkan, Metal, WebGPU, Apple Core ML and Snapdragon. It is also in Hugging Face Transformers from version 5.21 as the reference PyTorch implementation, which targets batched GPU throughput, for example to build speech datasets. The post lists demos such as a voice running on an Apple Watch with Wi-Fi off, an iPhone rep-counting app, a browser radio show and community conversions to ExecuTorch and Node.js. The weights ship in q8 GGUF form, a web demo lets anyone listen inside a browser tab with the text staying on the device, and the models let users set the speaking speed and normalize loudness.

Gradium says it announced Phonon in April and spent the next six months in private beta with partners, and that the feedback led to smaller models, more languages and support for custom voices. The authors add that Phonon uses flow matching, which differs from the techniques in most other TTS systems, so training was numerically unstable at first and needed several tricks to give good quality in all five languages. In Transformers the model generates audio latents one step at a time, each output fed back as the next input, and the audio codec plays the role that a tokenizer plays for text. The Transformers version is described as about 200 lines of code that list only what differs from Pocket TTS.

The limits are plain. Only the 31m English model is under a permissive license. The accuracy test uses hard structured sentences and a language-model scorer, so it measures that kind of text, not speech in general. Speed is measured on a few devices, with Kokoro's streaming difference noted. The naturalness result is a listening panel run by the vendor. The post says Gradium trains custom models with design partners, for example for Korean, Arabic or Japanese, which means the released models do not cover those languages. We did not run the models or reproduce any figure.

Our reading: small, fast, offline TTS is useful for privacy, cost and latency, and this post gives developers a clear list of sizes, licenses and platforms. Developers building commercial apps should note which model carries which license. Treat the ranking as the Gradium/Hugging Face coauthors’ leaderboard reading and performance numbers as Gradium’s tests; check both on your own text, voices and devices.