Second the recommendation for Snips. I recently built a Snips-based voice assistant for hands-free oscilloscope control (https://github.com/jmwilson/ollie), and snips is a much better fit for this kind of domain-specific, field-deployable (offline) voice assistant.

Consider the whole of infrastructure required for this application. The Pi sends a request to Alexa, which sends a request to a lambda server, which sends a request to SQS, which is polled by the Pi. Each of those AWS servers has possible subsystems and some load balancer in front of it. Single device, offline processing is so much cleaner when there's no need for general connectivity.