livekit-wakeword
livekit-wakeword is an open-source wake word library developed by LiveKit, released in February 2026. It uses a conv-attention neural architecture that processes mel spectrograms through frozen Google speech embedding and openWakeWord embedding models, producing a feature matrix that is temporally aware — detecting the order of speech events rather than just their co-occurrence, which substantially reduces false triggers from phonetically similar phrases.
Its headline accuracy claim is compelling: 60x lower false alarm rate and 100x fewer false positives per hour compared to openWakeWord, with 86.1% recall on English validation data. These figures represent a meaningful improvement in false-positive handling — a critical metric for a distress app where every false trigger could dispatch armed response vehicles unnecessarily.
The library supports custom wake word training for 30+ languages via VoxCPM2 TTS synthetic data generation, and the Apache 2.0 licence permits commercial deployment. However, livekit-wakeword has two critical limitations for the proposed SA distress app. First, it provides no Android SDK — only Python, Rust, and Swift (iOS) are supported. Given South Africa’s estimated 85%+ Android market penetration, an iOS-only or iOS-primary detection engine is commercially impractical. Second, the project is brand new (v0.1.0, February 2026) with only 105 GitHub stars, meaning it lacks the production validation and community support of more mature alternatives.
Connections
- openWakeWord — competes_with (same architecture base, lower FPR), source: https://github.com/livekit/livekit-wakeword
- Picovoice Porcupine — competes_with, source: https://github.com/livekit/livekit-wakeword
- iOS Background Audio Constraints — limited_by (iOS-only deployment), source: https://github.com/livekit/livekit-wakeword
Ontology livekit-wakeword [competes_with] openWakeWord livekit-wakeword [competes_with] Picovoice Porcupine livekit-wakeword [relates] iOS Background Audio Constraints