Gradium’s October 20, 2026 release of downloadable Phonon models gives people building read-aloud, accessibility and interactive apps a way to generate speech on a device. Once the necessary files are in place, the speech component can work without sending each passage to a cloud service, according to its runtime documentation.
Phonon 31m permits commercial use; the multilingual 40m model needs separate commercial terms. That distinction makes language and licensing the first decisions—not which model has the most impressive speed claim.
Choose language and license together
Option | Documented languages | Commercial-use position |
|---|---|---|
English only | Apache 2.0 permits commercial use, subject to its terms. | |
Separate models for English, French, German, Spanish or Portuguese | Published CC BY-NC terms exclude commercial use; arrange a separate license with Gradium. | |
English, French, German, Spanish and Portuguese in one model | Published CC BY-NC 4.0 terms exclude commercial use; arrange a separate license with Gradium. |
For an English commercial app, 31m is therefore a sensible first candidate to evaluate. For several supported languages, the multilingual model avoids selecting a separate model for each language, but its downloadable files do not carry the same commercial permission.
The 31m model card applies Apache 2.0 to the weights, tokenizer and voices in that repository—not automatically to voices obtained elsewhere. Apache’s redistribution conditions include providing the license, retaining required notices and identifying modified files. The CC BY-NC license requires attribution as well as restricting commercial use.
Gradium directs commercial 40m enquiries to sales; its comparison page provides no numeric price. Confirm fees and distribution rights before budgeting for a product that depends on those models.
What must be ready before going offline
Offline operation starts with preparation. The runtime guide documents first-use downloads for several integrations and loading from a local model folder without Hugging Face access. Have the runtime, model files and chosen voices available before connectivity disappears. The 40m downloads require acceptance of their access terms.
For web apps, the browser package documentation says a previously loaded model can reload offline from its cache. Voices not loaded earlier are fetched on first use. Persistent caching needs HTTPS or localhost, and the browser must support WebAssembly SIMD and module workers. A 40m deployment needs authorized downloads served with the app, rather than direct loading from the gated repository. The app’s own interface must also remain available offline.
The model cards list compressed, 8-bit weight files of 64.5 MB for 31m English and 74.6 MB for 40m multilingual. Those are file sizes, not total app storage or working-memory requirements; budget separately for the runtime and other assets.
The practical distinction is between offline speech and an offline product. An app that still sends text to a cloud assistant for a reply has not become fully offline merely because it reads that reply aloud locally.
Long passages and numbers need different checks
The model-card limitations say very long inputs are divided into sentences, so prosody—the rhythm and intonation of speech—does not continue across the whole passage. Reading an entire document is therefore not the same as shaping its delivery as one continuous narration.
For a short instruction, assess whether the words are clear and speech begins promptly. For an audiobook or extended lesson, assess transitions and emphasis across several paragraphs. The documented limitation is a reason to evaluate sustained narration separately, not evidence that every long passage will sound poor.
Text preparation matters too. The Hugging Face Transformers integration does not spell out language-dependent numbers, money amounts or addresses. Its documentation recommends Gradium’s ptts text normalizer or writing those expressions out. For an app reading prices or contact details, check the actual input-processing path; using the same model does not guarantee identical pronunciation across integrations.
Gradium publishes device-speed benchmarks, but its release-page timing notes exclude the one-off work of preparing the selected voice. Those vendor results do not establish startup delay, battery use or memory needs on every user’s device. Include those factors in an evaluation before committing to a device range.
A custom voice is a separate decision
The documented custom-voice workflow creates or clones a voice through Gradium Studio or its API, then exports a voice file for local speech generation. Creating that voice is an online service step. Gradium also requires the speaker’s consent for cloning.
The Transformers documentation explains why downloading the model alone is insufficient: the component that turns a recording into a voice representation has not been released. Local speech generation does not mean local voice cloning.
Start with the model whose language coverage and license fit the product. Then evaluate the intended voices and real text on the target devices, including a disconnected restart. If the product needs multilingual commercial use, sustained narration or a newly created voice, resolve that specific dependency before promising users offline speech.