Spectral gating
Also known as: spectral noise reduction
Spectral gating removes noise by learning what the noise looks like across frequency, then muting bands that fall under a threshold. It is effective on steady hiss, but it can also mute the quiet edges of speech, which is where word endings get lost.
Why it matters
Spectral gating is the classic denoiser, and understanding it explains its classic failure. It learns the noise from a quiet moment and subtracts that shape everywhere. The problem is that the quiet end of a word, an unvoiced s or f or the soft decay of a vowel, looks a lot like noise, so the gate treats it as something to remove. The loud middle of a word survives; the fragile edges do not.
In practice
Use it gently and give it a clean noise sample to learn from. Always compare the cleaned file against the original on the same phrase, and if word endings start to disappear, reduce the strength rather than accepting the loss.