He wants to create a custom voice (aka Microsoft Sam or Siri's voice etc), audacity wont' cut it.
The immediate problem will be the quality of voicemail recordings (terrible) and the quantity of them - any AI software will need a large sample to build a proper profile for a custom voice. Usually these are done with specifically crafted sentences as well, as random collection of spoken words might not (very likely *will* not) cover all sounds/intonations necessary.
It's a very nice thought
but you'll need someone who knows ML to help you, I unfortunately won't have the time
"As a result of all this hardship, dirt, thirst, and wombats, you would expect Australians to be a sour lot. Instead, they are genial, jolly, cheerful, and always willing to share a kind word with a stranger, unless they are an American." -- Douglas Adams