Diagnose the sound before choosing the tool
Three different problems hide behind "the audio is bad". Background noise needs reduction, a low or uneven level needs loudness work, and wrong content — a licensed bed, a stale VO, an unwanted commentary — needs replacement or removal. Picking the tool before naming the problem is how noise reduction eats a voice, or a music swap happens where a five-decibel boost would have done. Listen first, name the problem, then act.
Clean first, level second
Noise reduction and loudness interact. Normalizing a noisy track raises the noise floor along with the signal, and reducing noise after normalization changes the loudness picture again. The workable order is to strip hiss and rumble first, then set the level — and when repair needs close attention, extract the track, clean it, and use the replacement step to put the improved sound back against the untouched video.
Loudness is a delivery contract
Platforms measure perceived loudness in LUFS and adjust uploads that miss their target, which is why EBU R128 normalization exists — it speaks the platforms' language instead of guessing. Normalize when the destination has a loudness expectation, record which target was used, and verify the result by measurement. Reserve plain volume changes for quick relative fixes where no standard applies, because "louder" is not a delivery contract.
Swap deliberately, silence cleanly
Replacement is a surgical operation — the video stream stays as it is, the new audio takes the old track's place, and sync is the risk. Line the new track to a recognizable moment, then check the end of the video, because drift shows up over duration. Removal is simpler but deserves the same respect: the goal is a deliverable that plays genuinely silent in the target player, confirmed by playing the actual file rather than trusting the export log.