Is there such a thing that works the inverse of the noise removal algorithym? In noise removal you find a specific portion of the track representative of the noise profile that you wish to remove and then the noise removal plug in removes anything on the track that is similar to the noise profile. It seems to me that you just need to change the plug in program to discard everygthing but the represenative sound.
Voice is also only within a specific frequency range especially if the suggested plug in would request the gender of the people in the conversation and the recording mechanism of the recording (cell phone). There are effects plug ins that will modify sound to simulate an FM radio, AM radio, and a phone conversation although they may not necessarily work in reverse.
I think that it would be possible to do this with a little thinking on the correct algorithym. The success of removing he vocals may also depend on if there were other voices the same vector distance and direction as the desired voices. Every instrument has a unique set of harmonics, in addition to the main frequency, including the human voice. If audacity could remove enough noise then the listener’s brain could do the rest. The proposed plug in program would have to allow the user multiple inputs to allow human interaction to decide what to keep and what to filter.