Apple AI Updates: July 29, 2026

1. Apple Details The On-Device Model Behind Siri Expressive Voices

Apple researchers published work on memory-efficient audio synthesis using decoupled temporal and depth diffusion transformers, the architecture behind Siri Expressive Voices. The system runs entirely on-device on Apple’s Matrix Coprocessor, generating streaming speech at roughly 10ms per step, about 16x faster than real time, while using only around 21MB of peak runtime memory through causal sliding window attention with fixed-window key-value caching. The approach lifts Mean Opinion Score by 0.28 over the previous on-device system and supports continuous synthesis of 20 to 320 seconds with pace and expressivity controls, showing developers a path to high-quality real-time TTS under tight memory budgets. Source