Speech Enhancement

Older noise reduction subtracted what it thought was noise and left the voice thin and underwater. Enhancement tools work the other way round, using a model trained on clean speech to resynthesise the voice, which is why they can strip out a reverberant room in a way a gate or a denoiser never could.

The limits show up on anything the model has to guess at. Heavy processing makes voices sound slightly synthetic and oddly similar to each other, background music and laughter get mangled, hard sibilance can smear, and clipped audio is gone for good. Record the best source you can and use enhancement to tidy it.