Most AI voices live in the cloud: every sentence travels to a server and back, which costs money, adds delay and sends your text somewhere else. Gradium's Phonon takes the opposite route. Its text-to-speech models are small enough to run on a phone, a laptop, a single-board computer or a browser tab with no network at all, and two weeks after release the company has set out how they have fared.
There are two sizes, at about 65 MB and 75 MB. Phonon 31m speaks English and is released under Apache 2.0, so it can go into commercial products free of charge. Phonon 40m adds French, German, Spanish and Portuguese under a non-commercial CC BY-NC licence; commercial use of that model is priced by number of users rather than by how much speech you generate. There are more than 100 built-in voices, custom voices can be cloned or designed in Gradium Studio, and the models are meant to handle phone numbers, dates and codes that trip up many systems. Gradium says Phonon was downloaded more than 200,000 times in its first two weeks.
The benchmark result is the newsworthy part. On Hugging Face's new Open TTS Leaderboard, which scores models in 23 languages on intelligibility, speed and naturalness, Phonon ranks first for intelligibility in German and Spanish and in the top three in every language Gradium evaluated, alongside models up to 100 times larger. It also tops the CPU streaming benchmark on time to first audio, the delay before a voice starts speaking. Gradium built its own inference framework, ptts, with more than 10 back ends across CPUs, GPUs and phone chips, and the model is in Hugging Face Transformers from version 5.21.
Our Gradium Phonon review already rated the 31M English model as the tiny voice to ship in an app today, smaller than Kokoro and preferred by listeners. On-device speech matters for apps that cannot send text to the cloud, such as health, education or offline tools, and for anyone who wants a voice that keeps working without signal; demos include an Apple Watch talking through a coffee brew with Wi-Fi off.
Some caution is due. The leaderboard positions come from Gradium's own write-up, and intelligibility measured by word error rate says little about how natural a voice sounds over a long passage. The multilingual model's non-commercial licence will rule it out for many businesses unless they pay, and languages beyond the five, such as Korean, Arabic and Japanese, are offered only as custom training for partners.