Perhaps it's possible to use some sort of active cancellation - like what noise-cancelling headphones do - only in this case you would be trying to minimize the correlation between gyroscopic data & microphone data.