You can use ML to optimize for the vocals e.g. like noise cancellation. The upside of this is that you will almost certainly get crystal clear voice output (since 1) you have the known music piece, may not be the exact copy but the signal is there 2) the voice is clear since the music is background music). The only downside is that clear voice is not the same as original voice and it may sound unnatural, but you won't have crackling or any of the other artefacts of traditional filter. This is purely in theory of course, I have not tested it yet.