Open-weight audio and on-device speech

Phonon TTS brings small offline speech weights with a sharp license split

Gradium's technical deep dive clarifies how tiny offline speech models work and why the commercially usable 31M release differs from the 40M family.

Open Model Weights published5 Nov 2026
Primary sourceGradium — Phonon TTS technical deep dive on Hugging Face
Source published2026-11-05
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original wide conceptual graphic titled PHONON TTS showing an abstract waveform and local speech-synthesis circuitry; not a documentary image or independent performance chart.
Original Open Model Weights conceptual editorial artwork; not a documentary image or independent performance measurement. AI-generated editorial illustration; not a documentary photograph.

Gradium's November 5 technical report takes a closer look at Phonon, its family of small on-device speech synthesis models. The weights were opened earlier, so today's story is about how the released models work, where they can run and what their results mean. One distinction matters more than a blanket claim of openness: the English 31-million-parameter version permits commercial redistribution under Apache 2.0, while the larger variants use a noncommercial model license. Primary source ↗

The official phonon-31m-en repository describes a 64.5 MB eight-bit GGUF model that streams English speech at 24 kHz. Its model weights, tokenizer and included voices are listed under Apache 2.0. Gradium also offers approximately 40-million-parameter versions for English, French, German, Spanish and Portuguese, together with a multilingual option. Those weights are licensed under Creative Commons Attribution-NonCommercial 4.0. Developers evaluating business uses must examine the precise model and voice files rather than assuming the commercial terms of Phonon 31M apply throughout the family.

The system follows the Pocket TTS lineage, predicting audio latents that are decoded into sound. Gradium maintains a native ptts runtime aimed at device-side execution and documents a Transformers implementation for developers who prefer that ecosystem. This makes local voice output interesting for private assistants, games, accessibility features and applications where sending each sentence to a remote server would add cost or latency. A local runtime still may need an initial checkpoint download; device memory, language normalization, voice assets and backend availability all matter.

Gradium reports strong intelligibility and first-audio latency results on the Hugging Face Open TTS Leaderboard, including German and Spanish language evaluations. Those are linked to the evaluation's own methodology and the publisher's tested configurations, not measurements independently repeated by Open Model Weights. The practical next step is to benchmark cold-start memory, real-time streaming, pronunciation and selected voices on the target device. License compliance and consent for voice cloning remain separate from any performance claim.

OMW REGISTRY WATCH

What this changes for the evidence layer

Phonon's November 5 article is a technical follow-up rather than a new checkpoint released today. The 31M Apache-2.0 and 40M CC BY-NC 4.0 records must be treated separately before any source-pinned registry admission.

PRIMARY SOURCE

Gradium — Phonon TTS technical deep dive on Hugging Face

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗