Long read · 8 minutes

Restoration, and why the flat file comes first

Every job leaves here with a transfer that has been shaped by exactly one decision - the equalisation curve - and nothing else. Whether a second, cleaned file is worth paying for depends on what the recording is for, and that is a conversation, not a default.

The argument for flat

A flat transfer is the only file that can be re-derived. Ten years from now the tools will be better than ours; if we hand you a de-clicked, de-hummed, de-essed master and throw the raw capture away, that improvement is unavailable to you forever. Worse, restoration decisions are invisible after the fact: a notch at 120 Hz looks like a quiet recording, not like a filter.

So the deliverable we consider mandatory is the one nobody enjoys listening to. It has the hiss, the hum, the two seconds of leader, the engineer clearing his throat. It is the document.

What we will do, and what it costs you in sound

Click and crackle removal

On lacquer and shellac transfers this is the single most effective step and the one with the smallest penalty. We work in passes with a short detection window, aiming at impulses shorter than two milliseconds, and we do not let the algorithm run at a strength where cymbals or consonants start registering as clicks. The tell that someone went too far is a recording where the brass sounds like it is behind a curtain.

Hum and its harmonics

Mains hum at 60 Hz with harmonics at 120 and 180 Hz comes off cleanly when the pitch is stable. It is not stable on tape that was recorded on a machine with a slipping capstan, and chasing a drifting fundamental with fixed notches leaves a warble that is more distracting than the hum. In that case we say no and suggest the flat file with a gentle high-pass instead.

Broadband hiss

Hiss reduction is where most damage gets done. Our rule is a maximum of 6 dB of broadband reduction, measured on the noise floor between programme, and we will go further only with a written request naming the figure. Past roughly 10 dB, reverb tails start to pump and a solo voice develops a metallic edge that cannot be undone afterwards.

Dropouts and tape damage

A dropout from a crease is a hole in the signal, not noise, and nothing fills it honestly. Where the hole is under 40 milliseconds we interpolate and mark the timecode in the log. Longer than that we leave it and tell you where it is. We have never been thanked for inventing a quarter-second of music.

What we will not do

• Pitch and speed «correction» by ear. If the recording has a known reference - a test tone, mains hum of known frequency, a tuning note - we will correct to it and show the arithmetic. Without a reference, your memory of how a voice sounded is not a measurement, and we say so politely.
• Stereo widening of mono sources. It is irreversible and it corrupts the document.
• Loudness normalisation of preservation masters. Access copies, yes, to a stated target; masters stay at the level the tape gave us.
• Separating two overlapping programmes on a cassette recorded over twice. Source separation tools can make something that sounds impressive and is not what was on the tape.

The processing log

Every restored file ships with a text log in the same folder. It is written for a stranger:

item: 26-0431-007  side A  lacquer 12 in, 78 rpm
source file: 26-0431-007-A-flat-24-192.wav   (md5 in manifest)
1. resample 192 -> 96 kHz, linear phase
2. de-click: window 1.8 ms, strength 3 of 10, two passes
3. de-crackle: strength 2 of 10, one pass, 300 Hz - 8 kHz only
4. hum: notches 60/120/180 Hz, Q 24, -14 dB total
5. high-pass 28 Hz, 12 dB/octave  (rumble from worn centre hole)
6. no hiss reduction applied - operator judgement, programme too quiet
operator: KO   date: 2026-09-29   time on item: 1 h 10 min

We keep a copy of that log with the job for three years. If you come back later and ask why a file sounds the way it does, the answer exists in writing and is not somebody's recollection.

Two short cases

A county society sent us eleven oral-history cassettes recorded in 1991 on a portable deck in a kitchen. The tapes were fine; the refrigerator was not. We transferred flat, then offered a restored set with the compressor cycle notched out at the four frequencies it occupied. They took both and publish the restored set online while the flat masters sit on their shelf.

A guitarist brought a 1996 DAT with a mix he had been told was unusable. The tape read clean at 44.1 kHz except for nine uncorrectable blocks, all within one four-second span of a fade-out. No restoration was needed at all - the right answer was a careful transfer, an honest note, and an invoice for one tape rather than an hour of filtering.