Crosstalk 1.1 mini

In July we introduced Crosstalk 1 mini, our smaller, faster, cross-platform text-to-speech model that clones voices entirely on-device, on phones going back to 2020.

We have since reworked the engine underneath it for up to 4x better performance, which we're releasing as Crosstalk 1.1 mini.

On iOS, the model originally ran on the CPU and GPU. The core of it now runs on the Apple Neural Engine instead, which does this class of arithmetic faster and at a fraction of the energy.

Android phones have accelerators too, but it's tricky to take advantage of them effectively, so Crosstalk 1.1 mini still runs on the CPU on these devices. Even so, we found some significant performance gains, which have roughly halved the generation time.

Here's the before and after:

Against July's engine, that's 3.7 times faster on the iPhone 12 Pro and nearly twice as fast on both Android phones.

Crosstalk 1.1 mini will be rolling out in our flagship app within the next few weeks.