I have been trying to use custom keyword using the speech devices sdk but have had problems when I use my own custom keyword and deploy to android phone (the standard ones are better but still not as good as I need or would expect in commercial application). The screen shots on linked page implies that you can "Add training data to train keyword model" however that doesn't appear when I use the speech studio.
My suspicion is that the generated speech files that are automatically created by the speech studio are not good enough to train model for users with accents (like myself).
We have not yet widely enabled the KWS model adaption.
The Custom Keyword generated from the portal aims to be sufficient for initial trial, it is not currently at the level for commercial application. We are enabling the ability to upload data to adapt the model, this is being trialed with customers before wider roll-out. It is an upload on the Custom Keyword page and not the Custom Speech page.