Gradium has released Phonon, a small text-to-speech SI model designed to run entirely on user devices without requiring server connectivity or network access. The launch marks a shift for local voice SI synthesis, offering high-quality output that competes with larger cloud-based systems while keeping user data private on the device.
What Happened
Phonon is available in two sizes, approximately 65 MB and 75 MB, supporting English, French, German, Spanish, and Portuguese. The smaller model, Phonon 31m, is licensed under Apache 2.0 for commercial and non-commercial use, while the larger Phonon 40m is available under a non-commercial license. The models feature over 100 built-in voices and support voice cloning and design through Gradium Studio. The release builds on the Pocket TTS lineage, utilizing flow matching for audio generation, a technique distinct from many other TTS systems. Training the models involved addressing numerical instability to achieve consistent quality across all five supported languages.
The tool integrates with the Hugging Face Transformers library starting from version 5.21. Gradium also developed a new inference framework called ptts, which supports over 10 backends including CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU), and NPU (Apple Core ML, Snapdragon). This allows Phonon to run on a wide variety of hardware, from single-board computers to browsers and mobile devices. In its first two weeks, the model was downloaded more than 200,000 times.
Why It Matters
For developers and users concerned with data privacy and latency, Phonon represents a significant step in on-device SI capabilities. By eliminating the round trip to a server, the model ensures that text inputs remain on the user's device. This architecture is particularly relevant for applications where network connectivity is unreliable or non-existent, such as offline navigation or assistive technology on mobile devices. The integration into the Transformers ecosystem allows developers to leverage standard PyTorch workflows, with the model structured to handle audio latents similarly to how language models handle text tokens.
On the Hugging Face Open TTS Leaderboard, a new comparative tool for evaluating open TTS models, Phonon demonstrates strong performance. The company reports that the model ranks first in intelligibility for German and Spanish, and places in the top three for every language evaluated, competing against models up to 100 times its size. It also tops the CPU streaming benchmark for time-to-first audio, highlighting its efficiency for real-time applications on standard hardware.
The Bottom Line
Phonon offers a viable open-weight alternative for high-quality, private, and efficient on-device speech synthesis. With its Apache 2.0 licensed core model and strong benchmark results in intelligibility and speed, it provides a practical foundation for developers building local SI applications without the infrastructure costs or privacy concerns associated with cloud-based services.