openWakeWord
openWakeWord is an open-source audio wake word and phrase detection framework developed in Python, focused on performance and simplicity. It uses a shared feature extraction backbone model (Google’s TFHub speech embeddings under Apache-2.0) to convert mel-spectrogram audio inputs into speech embeddings, then routes these through small wake word classification heads trained on 100% synthetic speech data generated from text-to-speech models.
The framework achieves competitive accuracy: its “alexa” model reportedly outperforms Picovoice Porcupine on tested datasets, with target thresholds of less than 5% false-reject rate and fewer than 0.5 false-accepts per hour. Critically, because each additional wake word model shares the same feature extraction backbone, the marginal cost of adding new wake word models is low — a design choice suited to a distress phrase detector that needs to listen for multiple language variants of “help me” simultaneously.
However, openWakeWord has three critical limitations for the proposed application. First, its pre-trained models are licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0, prohibiting commercial use. A commercial distress app would need to train entirely new models using commercial-licence training data, requiring substantial ML infrastructure and expertise. Second, the framework is English-only — no support exists for Afrikaans, isiZulu, isiXhosa, or any other South African language, making it unsuitable without major custom development. Third, there is no official mobile SDK — the framework runs on server/desktop/Raspberry Pi hardware only. Porting to Android or iOS would require significant native development work.
Development activity has been low since the last release (v0.6.0, February 2024). For very low-power or embedded hardware use cases, the documentation itself recommends alternative projects.
Connections
- Android Background Voice Recognition — affected_by (no mobile SDK), source: https://github.com/dscripka/openWakeWord
- Picovoice Porcupine — competes_with (alternative wake word engine), source: https://github.com/dscripka/openWakeWord
Ontology openWakeWord [competes_with] Picovoice Porcupine openWakeWord [relates] Android Background Voice Recognition