Gradium releases compact AI speech models for offline apps, with a licensing split

Speech generation no longer needs a server connection for apps using Gradium’s newly released Phonon models. The company published the compact, open-weight family on October 20, 2026, targeting phones, laptops and browsers. But the release presents a choice: an English model licensed for commercial products, or larger multilingual options with noncommercial terms.

This is the public weight release of a model Gradium first announced in April and then tested in private beta. The company says feedback from partners and early users shaped smaller versions, broader language coverage and support for custom voices. Developers can now download the models rather than rely only on hosted speech generation.

English freedom, multilingual restrictions

Phonon comes in two sizes, measured by the number of model parameters, and three release options. Gradium lists the following sizes for its eight-bit quantized weights, a compact numerical format. The smallest option is also the only one released under Apache 2.0.

• Phonon 31m: 31 million parameters, 64 MB of quantized weights, English support and 37 built-in voices. Its Apache 2.0 license allows commercial applications.
• Phonon 40m: 40 million parameters and 74.6 MB of quantized weights, with separate models for English, French, German, Spanish and Portuguese. These carry CC BY-NC licenses.
• Phonon 40m multilingual: the same parameter count and published weight size, but all five languages in one model, with 102 built-in voices. It also uses CC BY-NC.

The licensing difference matters more than the small gap in download size for a commercial multilingual app. Gradium directs commercial users of the 40m models to contact the company. It positions those versions for greater expressivity and accuracy on difficult text, while the 31m model prioritizes a smaller footprint.

Smaller does not lose every comparison

Gradium’s English accuracy table gives the smallest model an unexpected advantage over its larger siblings. Phonon 31m records a 1.70% word error rate, compared with 2.11% for the multilingual 40m model and 2.22% for the language-specific 40m version. Kokoro scores 3.50%, while NeuTTS Nano scores 8.71%. Lower is better on this measure.

Those are company-published results on a deliberately difficult test, not a general verdict on speech quality. Gradium assembled 500 sentences per language from hard cases submitted by developers, including phone numbers, email addresses and postal codes. It says each system reads the same sentences using four female voices per language, although some competitors have fewer available voices.

The models use text normalization to handle structured text such as numbers and dates. Gradium has also published the hard-sentence dataset on Hugging Face. For naturalness, a separate dimension from getting words right, the company used blind pairwise listening tests across the supported languages, collecting 17,270 votes from 1,189 raters.

Speed comparisons need similar care. Gradium reports roughly seven milliseconds to first audio for Phonon 31m on a MacBook Air M5 CPU, versus 517 milliseconds for Kokoro. But Kokoro does not stream: its first audio arrives only when the whole sentence is ready. Phonon’s timing excludes the one-time voice-conditioning step.

Local playback, separate voice creation

Phonon generates audio in short pieces rather than waiting to finish an utterance. Its transformer processes text and speaker information, then produces a representation for each 80-millisecond audio frame. A smaller version of the Mimi decoder converts those representations into playable audio. All model sizes are distilled from a larger Phonon teacher model.

The deployment options extend beyond a browser demonstration. Gradium supplies backends for CPUs, graphics processors and mobile accelerators, including WebAssembly, WebGPU and Apple Core ML. Its product page says that once a model and voice are downloaded, speech generation needs no connection, and the text and audio stay on the device.

Custom voice creation follows a different path. Users clone a voice from a short recording made with the speaker’s consent, or describe a new voice in words, through Gradium Studio or its API. They then download the resulting voice file and load it into Phonon for local use.

For developers, the release includes inference code and setup instructions in Gradium’s GitHub repository, with integrations for Rust, Python, Swift and JavaScript/WebAssembly. The repository also includes a WebSocket server and an OpenAI-compatible server, providing routes into existing applications as well as new offline builds.