What we measured, what we ship, and what is still in training. Every number on this page is taken from the instrument itself: parameter counts read from the shipped ONNX graphs, timings from a warm CPU at batch one.
Four neural networks in sequence. Text is encoded; a duration model decides how long each sound lasts; a flow-matching field sculpts random noise into a mel-spectrogram in eight fixed Euler steps; and a vocoder turns that spectrogram into a 44.1 kHz waveform. The field is the only iterative stage; everything else runs once.
Built on open foundations: an open flow-matching speech model with MIT-licensed inference code and Open RAIL-M weights, which we export to ONNX, optimize, and serve. Our contribution is the inference and serving stack; attribution is retained per those licenses. The remaining ~18M parameters are a style encoder used offline to build voices, never run at synthesis time.
Most diffusion systems need tens to hundreds of network passes per utterance: roughly ~1,000 for high-quality diffusion, ~50 for fast variants. Because the flow path is nearly straight, eight suffice. That, plus a network small enough to keep its weights in cache, is the whole trick.
Not an illustration: a real utterance from the instrument, its spectrogram plotted star for star. Hear every voice in the Atlas, or synthesise your own in the Studio.
An RTF of 0.18 on a commodity core is the whole business case. A GPU serving realtime voice runs at batch one, its worst case, and idles most of its silicon; a CPU at batch one is doing exactly the work it is good at. The floor that collapses is the price: $1.90 per million characters, 0.19¢ per 1,000, and it holds in every region, on-premise, and air-gapped.
At list prices: 9× cheaper than Polly · 21× cheaper than Cartesia · 53× cheaper than ElevenLabs. Per minute of audio: $0.0017, 15 paise; per hour, $0.10. 10,000 characters free, renewed monthly. Full detail on the pricing page.
The lab ships checkpoints, not promises. One instrument is serving; its successor is on the bench, trained with the same discipline: distil a large open teacher into a compact student, keep the fidelity, and never let a GPU into the serving path.