Cover: AI-generated editorial illustration.
On November 5, 2026, Gradium published a detailed look at Phonon TTS, the on-device speech model they soft-launched roughly two weeks earlier. The write-up covers architecture, benchmark results, and the inference stack built to support it.
What Gradium is claiming
According to Gradium, Phonon comes in two model sizes: roughly 65 MB and 75 MB. Together they cover English, French, German, Spanish, and Portuguese, with more than 100 built-in voices and support for voice cloning. The company says Phonon runs on phones, laptops, single-board computers, and directly in a browser tab, requiring no network connection once downloaded.
On the Hugging Face Open TTS Leaderboard, Gradium reports Phonon ranks first in German and Spanish intelligibility, and places in the top three for every language evaluated, competing against models up to 100 times larger. CPU streaming speed, measured as time-to-first audio, is where the company says Phonon leads most clearly.
Licensing and pricing
Phonon 31m (English-only) is Apache 2.0, making it free for commercial use. The 40m multilingual model ships under CC BY-NC. Gradium states that commercial licensing for the larger model is priced per number of users, not per character or audio minute generated, which is a notable departure from typical TTS billing.
Developer access
Phonon is available through Hugging Face Transformers from v5.21 using the GradiumPhononTTSForConditionalGeneration class. The model follows a standard chat-template prompt layout, so integration code stays consistent across checkpoints. A dedicated inference library called ptts supports more than ten hardware backends including Apple Core ML, CUDA, WebGPU, and Snapdragon NPU.
Why this matters in practice
Most production TTS pipelines send every string to a cloud endpoint, creating latency, cost, and a data exposure point. A 65 MB model that handles dates, phone numbers, and codes correctly, runs offline, and integrates into Transformers with standard tooling is a practical alternative for apps where privacy or offline reliability matter. The Apache-licensed 31m English model in particular removes licensing friction for commercial builders. If you are evaluating on-device voice for your product, models are browsable at opennan.com/models.