New phonon-40m-en text-to-speech model released on HuggingFace

On October 19, 2026, the model gradium/phonon-40m-en was published on HuggingFace with the text-to-speech task. The model has 51 million parameters, the weights take up 0.2 GB in the repository, and the largest file is model.safetensors at 0.1 GB. The safetensors and gguf formats are available. The license is cc-by-nc-4.0, and the code is released as open source. Details are on the model page.

What it means

Our audio generators section already has several text-to-speech models, and phonon-40m-en stands out from them noticeably in size. XTTS has a largest weight file of 1.7 GB, Fish Speech has 1.2 GB, and ChatTTS has 0.8 GB. Here, the entire repository weighs 0.2 GB, which makes the model noticeably lighter for local runs. For licensing, phonon-40m-en uses cc-by-nc-4.0 — the same as the ChatTTS weights. This permits use and distribution with attribution, but for non-commercial purposes only. XTTS is restricted by its own Coqui Public Model License 1.0.0, which also prohibits commercial use, while Fish Speech requires a separate agreement for any commercial use. The phonon-40m-en developer does not specify VRAM requirements. For comparison, the ChatTTS authors state a minimum of 4 GB of GPU memory for a 30-second clip. The availability of the gguf format for phonon-40m-en may simplify running on your own hardware, but there are no specific platform instructions on the model page. The model expands the phonon lineup: phonon-31m-en was released in the same days. These are early versions so far, and there is no data on synthesis quality or supported languages other than English in the name.