CH-02 · Restoration

Noise reduction algorithms

Restoration is estimation under uncertainty. Every method here guesses what the recording would have sounded like without the unwanted signal, and each guess carries an expense. The engineer's job is to keep that expense below the threshold where a listener notices the cure.

Capturing a noise profile

Almost every spectral method begins with a profile of the unwanted signal. The engineer selects a passage that contains no wanted sound — a moment of room tone, a pause between words, the lead-in before the first note — and the tool averages its magnitude across frames to produce a per-bin estimate. Averaging matters: a single frame is a noisy guess, while a few hundred frames converge on a stable curve. The profile is only as good as the passage chosen, so a cough inside the selection will be treated as part of the noise and subtracted from everything that follows.

A profile is tied to a recording, a microphone position and a gain setting. Reusing one across a different session is a common error. If the room changed, or the preamp level moved, the profile must be captured again. The profile also fixes the ceiling of what can be achieved: where the profile is high, more reduction is possible; where it is low, the algorithm has little room to work before it begins eating the wanted signal.

Spectral subtraction

The classic method measures the magnitude of each bin in each frame, subtracts the corresponding profile value, and leaves the remainder as the estimate of wanted signal. Before the result is returned to the time domain, the magnitude is recombined with the original phase of the frame, because the transform of a real block discards phase and the ear is far less tolerant of phase error than of magnitude error.

An over-subtraction factor exists for a reason: plain subtraction removes the mean of the noise but leaves a residual flutter below the original level, heard as a faint burbling between words. Raising the exponent and adding a small floor reduces that flutter at the expense of a slightly duller result. There is no setting that removes noise entirely without removing something wanted; the useful question is only how much of each the engineer is willing to trade.

Gates and downward expansion

A gate compares the envelope of the signal against a threshold and fades out anything below it. A downward expander is the gentler cousin: instead of muting, it attenuates below the threshold by a chosen ratio, so quiet passages fall away smoothly instead of switching off. Expansion is usually the safer choice for speech, because a hard gate on a quiet syllable produces an audible hole and a click at the moment of closing.

Three controls decide the result. The threshold sets where work begins. The ratio sets how steeply the signal is pushed down below it. The hold and release set how long the gate waits before it closes and how quickly it fades, and a release that is too short chatters on every breath. Setting the threshold by eye on a meter is unreliable; it is better to watch the gain-reduction readout while a representative passage plays, and to move the threshold until quiet room tone is attenuated but every wanted consonant survives.

De-hum and narrow-band removal

Mains hum and its harmonics are the easiest problems on this desk because they are perfectly predictable. A narrow notch at the fundamental and at each harmonic removes them with very little effect on the rest of the band, provided the notch is narrow enough and the frequency is exact. A harmonic comb is often better than a single deep notch, because hum rarely arrives alone. Where the mains frequency is stable, a tracking filter follows slow drift; where it is not, a bank of fixed narrow notches is more transparent.