local-wake
local-wake is a GitHub project that implements wake word detection using Dynamic Time Warping (DTW) on user-provided reference audio samples, rather than a trained neural model. Users record a few examples of their wake phrase, and the system detects future occurrences by comparing incoming audio against these reference samples using DTW sequence matching combined with pretrained Google speech embeddings (ONNX format).
This approach is relevant to the SA distress app because it bypasses the SA language model gap entirely — any phrase in any language (Afrikaans “Help Asseblief”, Zulu “Ngisiza”, etc.) can be used as a wake phrase without training a language-specific model. Claimed accuracy is 98.6% on clean same-speaker audio, but real-world accuracy on stressed/noisy speech in an emergency scenario has not been evaluated.
Key limitation: accuracy degrades with speaker variation, background noise, and acoustic mismatch between reference samples and live detection — conditions that are all likely during an actual distress event.
Connections
- South African Language Wake Word Gap — mitigates, source: https://github.com/st-matskevich/local-wake
- Picovoice Porcupine — competes_with (DTW vs neural network approach), source: https://github.com/st-matskevich/local-wake