The nyumaya audio recognition engine features the two wakewords Marvin and Sheila. They are highly accurate and run even on the small PiZero, using 20% CPU.
The hotwords have specifically been trained to cope with a moderate amount of background noise.
Depending on your target hardware you can use different models, which are a trade-off between accuracy and CPU usage.
It's a standalone application with no dependencies and can be installed with one git clone command. Check out our Github Repository.