Hugging Face has released a model for local speech synthesis, Phonon TTS
The Phonon TTS speech synthesis model runs directly on the device without internet or cloud. The smaller Phonon 31m version (65 MB, English) has the Apache 2.0 license, usable commercially as well, while the Phonon 40m version (75 MB, 5 languages) with voice cloning is under CC BY-NC.
A speech synthesis (TTS) model called Phonon has been published on the Hugging Face Hub, which, according to the source, runs directly on the user's device without needing a server or network connection — meaning the entered text stays only on that device. The model comes in two sizes: Phonon 31m (approximately 65 MB, English only) is released under the Apache 2.0 license, making it freely usable even in commercial products. Phonon 40m (approximately 75 MB, five languages — English, French, German, Spanish, Portuguese) is under the CC BY-NC license, meaning free for non-commercial use; for commercial deployment, according to the source, the price is based on the number of users rather than the amount of generated speech.
The model uses a flow matching architecture and builds on the Pocket TTS line of models, which, according to the source, differs from the approach of most other TTS systems. Training was initially numerically unstable and required adjustments to work well across all five languages. The model includes more than 100 built-in voices and allows creating custom voices through cloning or voice design in the Gradium Studio tool. It runs on phones, laptops, single-board computers, and in the browser.
In the first two weeks after release, the model was downloaded more than 200 000 times and, according to the source, it still remains among the trending models on Hugging Face. On the new Open TTS Leaderboard, which evaluates models from more than 9000 TTS models on the Hub based on intelligibility, speed, and naturalness of sound across 23 languages, Phonon, according to the source, ranked first in intelligibility for German and Spanish and in the top three across all evaluated languages, compared to models up to a hundred times larger. The source also states that the model leads the benchmark for first audio output speed on CPU.
A separate inference framework called ptts was created for the model, with support for more than 10 compute backends (CPU, GPU, NPU) and integration into the Transformers library starting from version 5.21, intended for batch processing and higher throughput on GPU. Further details about the technical implementation are not available in the source.
Why it matters
The model enables running speech synthesis completely offline and without sending text to a server, which is relevant for applications with privacy requirements or operation without an internet connection. The smaller variant with the Apache 2.0 license is freely usable even commercially without licensing fees, while for the multilingual variant, the commercial price, according to the source, is tied to the number of users rather than the volume of generated speech — this is a different economic model than cloud TTS services billed by the amount of audio.
Two audiences, two different impacts
What this means
For individuals
Developers and app creators can try local speech synthesis without an internet connection, free of charge even for commercial projects with the smaller model variant.
For a business
Companies developing products with voice output can deploy speech synthesis without needing cloud infrastructure and with a pricing model based on the number of users rather than the volume of generated speech, which affects the cost of running voice features.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.