Phonon: Tiny Text-to-Speech Model Beats Rivals 100x Its Size

AI company Gradium announced the release of its text-to-speech (TTS) model, Phonon, built in collaboration with Hugging Face, the AI model repository site. The model can be deployed locally on mobile devices or laptops, with no internet connection required. It comes in two sizes. The smallest version has 31 million parameters, takes up about 65 MB, and generates audio in 28 milliseconds. That compares favorably to another TTS model, Pocket TTS, which takes 165 milliseconds to generate audio. On speech clarity, Phonon ranked in the top three on the Hugging Face Open TTS Leaderboard, a public ranking of TTS models. It got there against models up to 100 times larger. The smaller model is released under the Apache 2.0 license, which allows commercial use. The larger model is under a non-commercial license.

Models like this are a great fit for adding voice features to an app. Think text-to-speech narration, voice feedback in a user interface, or accessibility features. Because the model runs locally, you don't have to pay a cloud TTS service per request. And since the smaller model uses the Apache 2.0 license, it's easy to drop into a commercial app without worrying about licensing issues. One caveat: the leaderboard measured clarity, not naturalness or expressiveness. So if you're thinking of swapping an expensive cloud voice solution for Phonon in an app where tone of voice matters, listen to the sample recordings first.