Small, focused libraries for voice activity detection, audio I/O, and macOS tooling. Most are extracted from past speech SDK work experience, refined into reusable open-source components.
fgvad
Smart VAD with dynamic endpointing, on top of ten-vad. Knows when speakers actually stop talking, not just when sound stops. 5.9% force-cut rate vs 87% on a real 25-minute lecture when the dynamic tail curve is disabled.
Rust core + Swift / Kotlin wrappers. iOS 16+ / macOS 13+ / Android API 26+. SPM + CocoaPods + JitPack — one-line integration on each platform.
WindWhisper
Offline voice input for macOS. Status bar app — click to record (up to 60s), click again to stop, transcript pasted to cursor. SenseVoice int8 model, Chinese / English / Japanese / Korean / Cantonese auto-detect, all offline.
micvol
macOS hardware-level microphone input volume control via CoreAudio HAL. Maximize input gain at the A/D conversion stage instead of digital AGC. Cargo crate + static library + CocoaPods.
Previously: 5+ years on a wearable voice SDK. Earlier: speech recognition SDK work. The boundary between VAD and ASR — endpoint detection strategy, dynamic thresholds, audio gain at the analog / digital interface — is where these libraries come from.