To activate the service, the first voice in the list of available voices is taken and used to run a test. In my case this is the voice "Aoede". The error that comes from the Google API is the following:
This voice currently does not support SSML input. Please try again with text only input.
I can't find anything in the documentation about voices that do not support SSML input. If one could discover from the API whether a voice supports SSML, we could filter out the voices that don't.