Memory Efficient Audio Synthesis with Decoupled Temporal Depth Diffusion Transformers
Apple's Siri just got a memory-efficient audio synthesis engine. On-device. Real-time. No cloud.
Why it matters
Apple is shipping production-grade on-device audio synthesis via Siri Expressive Voices, powered by its AFM 3 Core Advanced foundation model. This demonstrates a significant capability advancement in edge AI—converting semantic tokens to high-fidelity audio within tight hardware constraints (Apple Matrix Coprocessor), a technical and product milestone for voice AI at scale.
The key facts
6 to knowSiri Expressive Voices ship with on-device real-time audio synthesis
Powered by AFM 3 Core Advanced foundation model (Apple's most powerful on-device model)
Architecture: semantic audio tokens → residual vector quantization (RVQ) detokenizer
Runs on Apple Matrix Coprocessor (AMX) with tight compute and memory constraints
Three-component design for memory efficiency
Published as research on Apple ML blog (Jul 2026)
Go to the source
Apple Machine Learningmachinelearning.apple.com
Publisher excerpt: Siri Expressive Voices synthesize rich, configurable speech in real time and entirely on device, powered by AFM 3 Core Advanced, Apple’s most powerful on-device foundation model. This work presents the memory-efficient audio synthesis architecture behind that capability: a detokenizer that converts…