The nature of ML is to use a large corpus of examples (training set) in order to provide the probabilistic values in order to predict the current case.
The common approach is to collect the current results and add it to the training set; then use it to improve the accuracy of future predictions.
We will need to collect all the results to the children's interactions, both correct and incorrect in order to refine the approach. This will also provide a foundation for performance metrics.
I expect that the performance demands of ML would exceed the practical capabilities of the devices to do in any "real time" fashion... and this will also apply to the semantic analysis of speech recognition on the local device, without connecting to cloud based resources.