Vosk
Vosk is an offline speech recognition toolkit built on Kaldi, developed by Alpha Cephei. With 14,700 GitHub stars, it is the most widely-adopted open-source offline speech recognition library with mobile SDKs — more popular than both openWakeWord and Porcupine by star count. It provides continuous large-vocabulary transcription with small (50MB) models, zero-latency streaming, reconfigurable vocabulary, and speaker identification.
Vosk’s key strength for mobile deployment is its combination of genuine offline operation, genuine Android and iOS SDK support (Java/Kotlin AAR for Android, Swift for iOS), and Apache 2.0 licensing that permits commercial use without model licence restrictions. Its 50MB model footprint is manageable for mobile apps, and the reconfigurable vocabulary feature allows the recognition vocabulary to be restricted to a small phrase list — effectively turning Vosk into a keyword/phrase spotter rather than a full speech-to-text engine, which would be the appropriate mode for a “help me” distress phrase detector.
However, Vosk’s critical gap for the South African use case is its language coverage. Of 20+ supported languages, none are South African: Afrikaans, isiZulu, isiXhosa, Sesotho, and Setswana are all absent. A user shouting “Help Asseblief” (Afrikaans), “Ngisiza” (Zulu), or a mixed-language phrase would not be detected. Building a custom Vosk model for an SA language would require a substantial audio corpus in that language — which is now becoming more feasible given datasets like Swivuriso and OpenSLR 32 SA Languages, but still represents a significant ML engineering effort.
As a full ASR engine rather than a lightweight wake word detector, Vosk also imposes higher continuous CPU load than Porcupine. For always-on background detection, this translates to greater battery drain, making it less suited to the critical “continuous passive listening” requirement than a dedicated wake word engine.
Connections
- Picovoice Porcupine — competes_with (alternative approach: STT vs wake word), source: https://github.com/alphacep/vosk-api
- openWakeWord — competes_with, source: https://github.com/alphacep/vosk-api
- Android Background Voice Recognition — possible_component_for, source: https://alphacephei.com/vosk/android
- Swivuriso — possible_component_for (SA language model training data), source: https://arxiv.org/html/2512.02201v1
- OpenSLR 32 SA Languages — possible_component_for (SA TTS data for model training), source: https://www.openslr.org/32/
Ontology Vosk [competes_with] Picovoice Porcupine Vosk [competes_with] openWakeWord Vosk [relates] Android Background Voice Recognition Vosk [relates] Swivuriso Vosk [relates] OpenSLR 32 SA Languages