Audio6h ago
Apple Unveils Memory-Efficient Audio Synthesis Architecture
The AFM 3 Core Advanced model powers Siri Expressive Voices using a new detokenizer for on-device synthesis. This architecture converts semantic audio tokens into high-fidelity sound within the strict memory limits of the Apple Matrix Coprocessor. It utilizes a three-component design and residual vector quantization. This allows real-time, configurable speech without relying on cloud compute.