To reproduce:
Create a basic app that transcribes speech.
Add “Flubbery Dubbery” or a made up couple of words to a strings array and set it equal to the contextualStrings property of SFSpeechAudioBufferRecognitionRequest
For the recognition request being used, set the requiresOnDeviceRecognition Boolean property to true.
Transcribe audio and say the made up couple of words.
See that the device never correctly transcribes the made up couple of words.
Now set the requiresOnDeviceRecognition Boolean to false.
Transcribe audio and say the made up words.
See that the device correctly transcribes the made up words.
PS, I noticed that if you add a custom word as a contact in the Contacts app, then on-device recognition picks it up. So it seems it's possible, just not implemented quite right.