Launch of Phonon TTS Models
Published November 5, 2026 — Phonon introduces a series of compact text-to-speech models designed to operate entirely on user devices, ensuring that input text remains local and eliminating the need for server connectivity. The two primary variants released are Phonon 31m, available under an Apache 2.0 license for commercial use, and Phonon 40m, which is licensed under CC BY-NC for non-commercial projects.
Design and Language Support
These models are based on the Pocket TTS lineage and utilize flow matching to generate audio latents. The training process required stabilization techniques to manage numerical instability across five supported languages: English, French, German, Spanish, and Portuguese. Phonon is offered in two binary sizes, approximately 65 MB and 75 MB, and includes over 100 built-in voices.
Demos and Adoption
The launch featured several device-centric demos, including an Apple Watch guide for brewing coffee and an iPhone rep counter that function offline. Additional demonstrations included a talking robot named Lamp from Autonomous, browser-based applications like an offline dino runner, a sci-fi game, and a reading of Proust page by page. In its first two weeks, Phonon achieved over 200,000 downloads.
Performance on Hugging Face Open TTS Leaderboard
On the Hugging Face Open TTS Leaderboard, Phonon ranked first for intelligibility in both German and Spanish, and it placed within the top three across all evaluated languages. Notably, Phonon excelled in speed, topping CPU streaming benchmarks for time-to-first audio, outperforming models significantly larger in size.
Inference Framework and Integration
A new inference framework named ptts was developed to support a wide range of devices, featuring over 10 backends including CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), and NPU (Apple Core ML, Snapdragon®). Phonon is also integrated into Transformers from version 5.21 as a reference PyTorch implementation, showcasing approximately 30,788,354 parameters for the model and 10,305,313 for the audio tokenizer.
Voice Cloning and Custom Training
Voices are provided as precomputed speaker latents (gradium/phonon-voices) and can be cloned from short recordings with the speaker’s consent or designed in Gradium Studio. Gradium and its partners offer additional custom training options, including model size reduction, support for new languages, or vocabulary fine-tuning upon request.
Original source: Hugging Face Blog