I need a lightweight Android library that can continuously listen for a single wake-word in Spanish and trigger my own voice-assistant module when it is detected. The focus is on practical, real-time performance rather than research-grade perfection, so moderate accuracy—roughly in line with popular consumer phones set to Spanish—is acceptable. A model that sits inside an AAR or Maven package, starts with a simple API call, remains in low-power mode, and throws a callback when the phrase is recognised is exactly what I’m after. I’m flexible about the underlying engine: TensorFlow Lite Micro, PocketSphinx, Snowboy-style approaches, or a custom on-device neural net are all fine so long as they run fully offline and do not require cloud calls. The only hard requirement is that the recognition word be delivered in neutral Latin-American Spanish pronunciation and not default to English phonemes. Deliverables: • compiled library ready to drop into any Android Studio project (minSdk 23) • Kotlin or Java sample app showing start/stop, permission handling, and the callback in action • clear README explaining model training steps so I can swap in a new keyword later if needed If you already have a Spanish dataset or a proven keyword spotter, mention it; otherwise, outline how you’ll collect or synthesise training data. I’ll test by running the demo on several mid-range phones and gauging latency, wake-word hit rate, and false-trigger frequency in a typical living-room noise environment.