Pick one
A library of voices Grey learned from real recordings: actors, narrators and everyday speakers in six languages.
Grey
Grey
Leave a control on Auto and Grey picks what fits the voice.
How Grey turns text into speech, what it learned from, and how it compares.
Grey turns text into speech in two steps. A small text model works out the rhythm: how long each sound lasts. Then a flow-matching decoder paints the audio, starting from noise and refining it in a handful of steps, the way image generators work, but for 48 kHz sound.
It was trained from scratch in October 2026, version after version, on public speech datasets. Cloning needs no training at all: a voice encoder hears a few seconds of someone and steers the model toward them.
A voice keeps its own sound in every language, so a Dutch voice can read Japanese, with a little of its accent.
The same sentence, the same voice, every version. Grey went from its first word to today in seventeen days.
On Seed-TTS-eval, the standard test for cloning a stranger's voice, Grey's English is as clear as the real recordings. Matching the voice itself is where larger models are still ahead, and that is what the next version is built for.
| Model | Size | English errors | Mandarin errors |
|---|---|---|---|
| Real recordings | 2.14% | 1.26% | |
| F5-TTS | 336M | 1.83% | 1.56% |
| Grey v12 | 187M | 2.20% | 3.69% |
| CosyVoice 2 | ~0.5B | 2.57% | 1.45% |
| MaskGCT | ~1B | 2.62% | 2.27% |
Grey is a text-to-speech model built by one person on one graphics card, in the Netherlands, in October 2026. It speaks six languages, learns a new voice from a few short recordings, and lets you shape how it talks: calmer, faster, happier, or in a whisper.
A library of voices Grey learned from real recordings: actors, narrators and everyday speakers in six languages.
Move a few sliders through the space of all the voices Grey knows, and you get a new voice that belongs to nobody.
Read a few sentences aloud or upload recordings. Grey learns the voice in seconds, with no training.
Clone your own voice, or someone who gave you permission. You confirm this before every clone.
Voices you clone are private to your browser and deleted automatically after 24 hours. Your clips are only playable by you.
If you share what Grey made, say it's synthetic. That's the whole point of the name.