Gradium’s small text-to-speech model, Phonon, has placed first on a speed test run by Hugging Face. The company shared the results on November 5, 2026, two weeks after it opened the model to the public.
Phonon 31m posted a 95th-percentile time to first audio of 28 milliseconds on the CPU streaming benchmark. The larger Phonon 40m came second at 38 milliseconds, ahead of Pocket TTS at 165 milliseconds.
Time to first audio measures how fast a model starts speaking after it receives text. A lower number means a shorter wait for the listener.
How the Leaderboard Scored Phonon
The Open TTS Leaderboard is a new Hugging Face project. It gives developers one place to compare open models from the more than 9,000 text-to-speech models on the Hugging Face Hub.
Models are tested in 23 languages on intelligibility, speed and naturalness. Voice cloning models also get a speaker similarity score.
On intelligibility, measured by word error rate, Phonon ranked first in German and Spanish. It placed in the top three in every language Gradium evaluated, next to models up to 100 times its size.
The leaderboard also has a Listen tab. Visitors can compare outputs side by side and vote for the voices they prefer.
A Model Built to Stay Offline
Phonon runs on the user’s own device with no server and no network. Text typed into it stays on the device.
The models come in two sizes, about 65 MB and 75 MB. Together they speak English, French, German, Spanish and Portuguese, with more than 100 built-in voices.
Gradium said small speech models have often sounded poor, while good ones were too large for small devices. Phonon uses flow matching, a method from the Pocket TTS lineage, to address both problems.
The company said training was at first numerically unstable. It needed several tricks to get quality results in all five languages.
Phonon runs on phones, laptops, single-board computers and in a browser tab. Gradium also built a new inference framework, ptts, which supports more than 10 backends across CPU, GPU and NPU chips.
Licensing differs by model. Phonon 31m speaks English under the Apache 2.0 license and is free even for commercial products, while Phonon 40m covers all five languages under a non-commercial license.
For commercial use of the non-commercial models, pricing is based on the number of users. It is not based on how much speech a product generates.
Launch demos included an Apple Watch guiding a coffee brew with Wi-Fi off and an iPhone counting workout reps with no signal. A robot called Lamp, from Autonomous, also spoke in character.
The Hugging Face team built a Transformers integration, available from version 5.21. Outside developers have added an ExecuTorch browser version and Micdrop support for Node.js apps.
Gradium first announced Phonon in April and worked with partners for six months before release. It now offers custom models, including smaller versions and training for Korean, Arabic or Japanese.
Phonon was downloaded more than 200,000 times in its first two weeks. The company said the model is still trending on Hugging Face.
