Release notes

0.1.0 · 14 November 2026

The first release includes the Android AAR, Swift integration, C ABI v1 and a native WAV-to-JSON example. It reports time-stamped sound intervals, peak model scores, detected sound duration, clip-eligible duration and grouped episodes. The model and detection settings ship with the SDK and inference runs offline.

Detector parity

Thirteen detector and model files match the active iOS and Android app sources byte-for-byte. Both mobile C cores passed all sixteen of their host tests. SDK output matched the reference C detector’s timestamps, durations and peak scores at 16, 44.1 and 48 kHz on the synthetic fixtures.

Sixteen consecutive 30-minute silent parts produced 28,800 recorded seconds and zero detected events on a Linux host. The tests also covered capture gaps, failed reads and retry.

Platform coverage

Android. A signed, R8-minified host app loaded the release AAR and ran inference on an API 26 x86_64 emulator. The release contains four ABIs. Physical phone validation is still open.

Apple. Swift 6.0.3 built the package and an independent SwiftPM consumer on Linux. The AVFoundation reader, XCFramework build and an iPhone run require validation on a Mac and an iPhone.

Native. The release CMake library, downstream C++ integration and WAV CLI were tested on Linux. Evaluate memory, power and throughput on your own target hardware.

Evaluate this release

Use the reference recording to verify timestamps and durations, then compare labelled recordings collected on the microphones you plan to support. The synthetic tests establish implementation parity. This release reports acoustic snore evidence and does not produce sleep stages or an apnea diagnosis.