Gradium releases Phonon for speech generation on the device
The downloadable model family turns text into speech on phones, laptops and browsers, with local generation designed to work without a network.
Gradium’s October 20 announcement introduces the downloadable Phonon family for text-to-speech: converting written input into spoken audio. The company built it around local generation on phones, laptops and browsers. For developers adding a speaking interface, the central change is that producing the voice need not involve sending each text to a remote speech service.
The English model card describes Phonon 31m as streaming speech while generation continues, with mono output at 24 kHz. The language setting controls normalization of numbers and dates; this model uses en. It is an English model: other languages can be mispronounced. Long inputs are divided into sentences, so their rhythm and intonation do not carry continuously across the entire text.
The runtime repository makes the setup boundary clearer. Its Rust implementation needs no PyTorch, and its uv quickstart writes a WAV file from text. The first run downloads and caches the default model. A requested voice missing from the checkpoint can trigger another download through the command-line or Python interface. Local checkpoint folders require no Hub access. Thus local speech generation should be distinguished from initially acquiring its files.
The October 22 MarkTechPost report also covers the release. Its benchmark table explicitly draws on Gradium’s blog and model card, rather than supplying independent measurements.
Sources(4)
- Gradium releaseChecked:
- English modelChecked:
- Runtime guideChecked:
- MarkTechPost reportChecked:
