Improving accuracy of industry terms

I am trying to develop streaming speech-to-text for medical use. While Amazon Transcribe Medical is good for most medical terminology, I cannot get it to recognize common dosing and rate phrases.

Examples of failing phrases:

1- "Administered at a rate of 5 mcg/kg/hr"

2- "begin at a starting dose of 8 mg/kg"

The spoken pronunciation of the dosing would be:

1- "mike-per-kig-per-hour"

2- "mig-per-kig"

In both of these examples, the fractions are either ignored or read out as "mike pir" or "months per cake", or sometimes it just crashes with a 500 error.

I have tried:

1- Adding a custom vocabulary with the following relevant entries:

Phrase	IPA	SoundsLike	DisplayAs
mig-per-kig		mig-per-kig	mg/kg
mike-per-kig-per-hour		mike-per-kig-per-hour	mcg/kg/hr

2- I tried varying the medical specialty with no improvement.

3- I tried using the regular Amazon Transcribe with the custom vocabulary, as well as a 10,000 word language model with many examples of these phrases in context.

4- I can get more reliable results by not verbally abbreviating the phrase. Instead of saying "mikes per kig per hour", saying "micrograms per kilogram per hour", but demanding providers to speak in this verbose manner is going to be met with annoyance and resistance to using the service.

Am I using the vocabulary table wrong? Is there something else I can do to improve the accuracy of these phrases? It pains me because the accuracy is otherwise very good, but missing common dosing phrases is unfortunately a small but critical deal breaker.