# Gradium details on-device speech models

> Gradium describes local speech licences.

By BIGCHANGE.AI Editorial

Published: 2026-11-05T14:47:38.639Z
Updated: 2026-11-05T14:47:38.639Z
Canonical: https://bigchange.ai/blog/gradium-phonon-on-device-speech

![Conceptual illustration of a portable speaker and phone on a plain surface.](https://bigchange.ai/api/media/file/phonon-local-speech-concept.png)
AI-generated conceptual illustration by BIG CHANGE.

Gradium’s [November 5 account](https://huggingface.co/blog/phonon-tts) describes Phonon speech models released about two weeks earlier.

## The big change

- After model and voice downloads, synthesis can run offline.
- 31M and 40M licences differ.
- Leaderboard results are metric-specific.

## Models and terms

The [31M English model](https://huggingface.co/gradium/phonon-31m-en) uses Apache 2.0, permitting commercial use under its terms; quantized weights are 64.5 MB. The [40M multilingual model](https://huggingface.co/gradium/phonon-40m-multilingual) has 74.6 MB quantized weights: English, French, German, Spanish and Portuguese. CC BY-NC 4.0 is noncommercial; access requires login and acceptance.

Single-language 40M variants share that licence; Gradium offers commercial terms. Counts exclude the audio tokenizer; file size is not runtime memory.

## From text to audio

The [Transformers documentation](https://huggingface.co/docs/transformers/main/model_doc/gradium_phonon_tts) describes a model that reads text and a voice prompt, generates audio representations incrementally and decodes them to speech. An application can receive audio before an utterance is complete. Gradium describes ptts for local streams and Transformers for batched GPU processing.

On-device synthesis can keep text local when the surrounding application is also designed that way. Gradium says its Studio or API can create a voice from a consenting speaker’s recording for local use. Voice creation is a separate service step. The Transformers implementation cannot clone directly from a recording: its documentation says the speaker codec is not released.

## What comparisons show

Gradium reports leading German and Spanish intelligibility and CPU first-audio latency on the Open TTS Leaderboard; BIG CHANGE has not tested it.

## Sources

- [Gradium](https://huggingface.co/blog/phonon-tts)
- [31M model](https://huggingface.co/gradium/phonon-31m-en)
- [40M model](https://huggingface.co/gradium/phonon-40m-multilingual)
- [Transformers: Phonon](https://huggingface.co/docs/transformers/main/model_doc/gradium_phonon_tts) — Inference and codec limitations.