Sub-20ms neural audio denoising on mobile

Media, Audio & Communications

Industry: Media, Audio & Communications

Platform: iOS

Stack: Swift · C · AVFoundation · Accelerate · SPM

Description: On-device neural noise suppression for real-time voice with end-to-end latency under a 25ms ceiling.

---

The engineering surface was systems-level audio work: package a C-only neural denoiser as a Swift Package Manager module without prebuilt binaries, replicate autoconf-generated defines in cSettings, wrap an opaque C state pointer with RAII semantics in a Swift class, and feed it from an AVAudioEngine tap with zero-copy buffer pointers. A heavier deep-learning alternative was built end-to-end and rejected against the latency budget; a CoreML spectrogram pipeline was prototyped and similarly rejected. Both rejections were documented with measured numbers so the call can be revisited as hardware shifts.

We owned the SPM module design, the C-source build, the Swift wrapper, the real-time pipeline, the latency evaluation of two heavier alternatives, and the XCFramework distribution. The client owned the VoIP product into which the module ships.

Engagement: single-sprint integration · single specialist · 2024.

Engineering signal: under 1ms inference per 480-sample frame; 15-20ms end-to-end pipeline latency under a 25ms ceiling; ~85KB binary footprint.