Gradium's open-weight text-to-speech, built to run on the device itself: phones, laptops, small boards and the browser. It comes as three models: Phonon 31m (64 MB), Phonon 40m (74.6 MB) and Phonon 40m multilingual (74.6 MB).

Does Phonon work offline?

Yes. Once the model and a voice are downloaded, Phonon needs no connection at all. Your text and the audio never leave the device.

Which languages does Phonon support?

English, French, German, Spanish and Portuguese. Phonon 31m speaks English. Phonon 40m comes as one model per language, and Phonon 40m multilingual speaks all five in a single model. Need another language? We build custom models on request.

Which devices does Phonon run on?

Any modern processor: laptops and desktops (x86-64 and ARM, Apple Silicon included), Android phones, even a Raspberry Pi 5. Apple's Neural Engine, through Core ML. Graphics cards, through CUDA, Metal or Vulkan. Qualcomm Snapdragon's NPU. The browser, through WebAssembly and WebGPU, in Chrome and Safari, on desktop and mobile.

Can I clone a voice or design a new one?

Yes. Make it in Gradium Studio or through the API: clone a voice from a short recording (with the speaker's consent), or describe one in words with Voice Design. Download the voice file, and Phonon runs it locally like any other voice. /voice-design.

Phonon: Open On-Device Text to Speech from Gradium

[ 1 ]

Introducing

Gradium's open-weight text-to-speech, small enough to ship in any app. Private, and it works offline.

Already trusted by

1.7%

Word Error Rate (WER)

Phonon 31m · English Tested on 500 sentences that other models get wrong: phone numbers, email addresses, postal codes.

Read the blog post

1.70%

Phonon 31m

2.11%

Phonon 40m multilingual

2.22%

Phonon 40m

3.50%

Kokoro

7.40%

Kitten TTS

8.71%

NeuTTS Nano

Phonon 1.0

Three models,
One for every device.

[ 2 ]

Phonon 31m

Phonon 40m

Phonon 40m multilingual

Languages

English

English, French, German, Spanish, Portuguese, one model per language

English, French, German, Spanish, Portuguese in a single model

Size

64 MB

74.6 MB

Parameters

31M

40M

Licence

Apache 2.0

CC BY-NCFor commercial use, talk to sales

Phonon 31m
Languages
English
Size
64 MB
Parameters
31M
Licence
Apache 2.0
Phonon 40m
Languages
English, French, German, Spanish, Portuguese, one model per language
Size
74.6 MB
Parameters
40M
Licence
CC BY-NCFor commercial use, talk to sales
Phonon 40m multilingual
Languages
English, French, German, Spanish, Portuguese in a single model
Size
74.6 MB
Parameters
40M
Licence
CC BY-NCFor commercial use, talk to sales

Where it runs

One model,
All of your hardware.

[ 3 ]

Any processor
Laptops, desktops, Android phones, even a Raspberry Pi 5. No graphics card needed.
iPhone, iPad and Mac
On Apple's own AI chip, through a Swift package. Up to 22× faster than real time on an iPhone 16 Pro.
In a web page
Runs in the visitor's browser, so their text never leaves their device. The demo at the top of this page is it.
On a graphics card
When there is one, for the fastest speech: NVIDIA, Apple, Vulkan or WebGPU, chosen when you build it.

Compatible with

No cloud,
just your hardware

Paste one of these prompts into your coding agent and it will read the README, build Phonon and make it speak. Or take the code.

View on GitHub Read the README

~/phonon/cli.prompt

1Set up Phonon, Gradium's on-device text-to-speech, from https://github.com/gradium-ai/xn-ptts.
2
3Read the README first and follow it over anything you assume.
4
51. Install the Rust CLI from the repository root with
6   cargo install --path ptts --locked --features cli.
72. Point MODEL_DIR at the model folder (config.json, model.q8.gguf,
8   tokenizer.json and the voices).
93. Synthesize "Hello world" with --lang en --quant q8 --voice Freya
10   to out.wav.
11
12Verify the file plays and tell me how to change the voice.

Text in, a WAV out. The fastest way to hear it.

Apps builtUsing Phonon

See all apps Submit your app

Harry Potter hat voice

Autonomous's robot, Lamp, puts on a sorting hat and talks in character, voiced by Phonon.

Demo

Magpie TTS meets Phonon

NVIDIA Magpie TTS meets Gradium Phonon: two open models in a scripted dialogue, Magpie in Adi's voice and Phonon in Sterling's.

Demo

Du coté de chez Swann

Proust's Du côté de chez Swann, read aloud page by page, with an Italian translation to read and hear alongside the French.

Try it

Xybrid game demo

A sci-fi video game on a desert planet, its voices generated by Phonon on the player's own machine.

Demo

Offline FM

An offline dino runner where Rex hosts a radio show, voiced live by Phonon inside your browser tab. Pick a voice, give Rex a topic.

Try it Demo

On watch

Phonon on an Apple Watch with Wi-Fi switched off, talking you through a coffee brew step by step, entirely on the wrist.

Demo

RepCoach

An iPhone app that counts your reps out loud and keeps you going, with no signal needed.

Demo

Chrome

Phonon Chrome extension

A Chrome extension that reads any text on the web aloud, with Phonon running in the browser.

Demo

Questions?We’re here to help

[ 4 ]

Start
Building

Explore GitHub Test it in the Studio

[ 1 ] Creazione di Adamo, by Michelangelo Buonarroti. Date: c. 1511, part of the ceiling painted 1508-1512. Material: fresco. Dimensions: 280 × 570 cm. Current location: Cappella Sistina, volta, Musei Vaticani, Vatican City.

[ 2 ] Les trois Grâces. Artist: unknown; Roman, with restorations by Nicolas Cordier and Amainci, 1609. Date: 2nd century AD (?). Material: marble, probably Greek. Dimensions: h. 123 cm, w. 86 cm, d. 29 cm. Current location: Musée du Louvre, Aile Sully, Salle 348, Paris. Inventory number: Ma 287; MR 211; N 642.

[ 3 ] Victoire de Samothrace. Artist: unknown, Rhodes (?). Date: c. 200-175 BC. Material: Parian marble for the figure, grey Lartos marble for the ship. Dimensions: h. 328 cm overall. Current location: Musée du Louvre, Aile Denon, Escalier Daru, Salle 703, Paris. Inventory number: Ma 2369; NIII 2447; MNB 2118; MNC 1433.

[ 4 ] Portrait de Juliette Récamier, née Bernard (1777-1849), by Jacques-Louis David. Date: 1800. Material: oil on canvas. Dimensions: h. 174 cm, w. 244 cm. Current location: Musée du Louvre, Aile Denon, Salle 702, Salle Daru, néoclassicisme, Paris. Inventory number: INV 3708; C 124.

Word Error Rate (WER)

Three models,One for every device.

One model,All of your hardware.

Any processor

iPhone, iPad and Mac

In a web page

On a graphics card

No cloud,just your hardware

Harry Potter hat voice

Phonon voice bot on Modal

Magpie TTS meets Phonon

Du coté de chez Swann

Xybrid game demo

Offline FM

On watch

RepCoach

Phonon Chrome extension

Questions?We’re here to help

StartBuilding

Three models,
One for every device.

One model,
All of your hardware.

No cloud,
just your hardware

Start
Building