voice identification accuracy improved 83% since 2018. they don't need to record everything - just enough to build voice signature and keyword profile