If you really want to do this thoroughly, you'd probably have to sample the same recording from different records several times. Then you can use a "consensus" algorithm, that reconstructs the original audio in some optimal way. (But better yet to publish the original recordings, so others can still try different algorithms).