WakeWord is an offline wake-word and keyword spotting plugin for Unreal Engine 5.6 and 5.7, powered by sherpa-onnx. Detection runs fully locally with low latency and no per-call cost, making it ideal for always-on voice triggers in games, interactive experiences, and AI dialog flows.

The plugin ships ready to use with bundled Chinese (wenetspeech) and English (gigaspeech) 3.3M models, including tokenizers, dictionaries, and Win64 runtime DLLs. No model download is required to get started. Add the Wake Word Component to an Actor, enable Auto Start On Begin Play, bind the On Wake Word Detected event, and you are listening.

Manage phrases from the Project Settings panel or at runtime through Blueprint and C++ APIs. Because each model is single-language, Chinese and English use separate phrase lists, and you can run both at once by adding two components.

Key features:

- Fully offline, low-latency keyword spotting with no per-call cost

- Bundled Chinese and English models, tokenizers, and Win64 runtime DLLs

- Blueprint and C++ APIs for adding, removing, and refreshing wake words at runtime

- Project Settings panel with per-word boost and threshold tuning

- Raw token line support and token validation against the model's tokens.txt

- Custom model support via the Custom preset for your own ONNX files

- Built-in events for detection, listening started, and error handling

- Dialog pipeline pattern to hand the microphone to a separate voice component

Compatibility and use cases:

- Engine: Unreal Engine 5.6+

- Platform: Win64

- Use cases: hands-free voice activation, NPC and assistant triggers, accessibility controls, and gating of downstream speech or dialog systems

https://github.com/zxwzxw/wakeword-unreal-plugin/blob/main/GitHub-Wiki.en.md