097 - Stash TTS Model

Stash, a read-it-later app that reads articles to you: https://juststash.app/

Part 97 of 100. Good text-to-speech models are usually rented. They run on someone else's server, cost money per word, and can disappear. For Stash we wanted voices that sound like a real person, running entirely on the phone, with no downloads and no dependency on anyone.

The catch: most TTS models run on the GPU, and iOS doesn't let apps use the GPU in the background. So we converted a model to Core ML to run on the CPU and Neural Engine instead. The whole model is about 130 MB, ships inside the app, and generates unlimited audio offline at zero cost.