Orrery — Copenhagen
The bench record.
Short entries, exact dates, the work only.
-
2026-09-08
On the bench: MD-1 · MiniDisc (ATRAC).
Yumeo MD-1 · ATRAC recorder · restored from unit no. M-2170534 · Yumeo, Japan, 1994
ATRAC, plainly — The mix runs dry, then through the recorder at SP from five seconds. What ATRAC discards at standard play is quiet — listen to the cymbal tails and the space around things, not the middle of the mix.
The instrument: https://orrery.dk/yumeo/md-1/
-
2026-09-06
Why a CD skips and stutters
A scratched compact disc does not fail all at once. It fails in stages, and each stage sounds different: first nothing at all, because the disc’s error correction absorbs damage that would ruin a simpler scheme; then a brief, almost undetectable smoothing where a sample or two goes missing; then a short, flat silence; and only at the end, the stutter — a repeated fragment or a jump — that most people mean when they say “the CD is skipping.” All four are real, distinct behaviours, and they escalate in that order as the damage gets worse: the first three belong to the disc format’s error-correction and concealment system, and the last is the transport giving up mechanically once that system has nothing left to offer.
The system doing the work is CIRC — Cross-Interleaved Reed-Solomon Code — the error correction built into the compact disc format from the start, part of the Red Book standard Philips and Sony published in 1980. CIRC does not try to protect any single stretch of the disc from harm. Instead it spreads each block of audio across a wide span of the disc surface before writing it, using two layered Reed-Solomon error-correcting codes with interleaving between them. A scratch that physically wipes out a short run of track — a burst error, severe by the standards of any single codeword — gets un-spread on playback into many small, separated errors scattered across different codewords, each recoverable from its own code’s redundancy provided the burst runs no longer than about 4,000 bits of track — roughly 2.5 millimetres. Interleaving is the whole trick: it turns one large, unrecoverable error into many small, recoverable ones.
When the damage is a deep gouge, a chip, a disc handled badly for years — enough to exceed roughly 4,000 bits of burst error — CIRC’s correction runs out, and the player falls back to concealment rather than correction. The first fallback is interpolation: estimate the missing samples from the genuine ones on either side, which for a short gap is inaudible, because most musical material changes little between adjacent samples at 44.1 kilohertz. Interpolation holds up to a much larger gap, roughly 12,000 bits — about 7.5 millimetres of track — before there is too little left to estimate from with any confidence; past that, the player mutes instead — a brief, deliberate silence, chosen because playing back whatever data actually survived would sound worse than playing nothing at all. And if the physical damage is severe enough to throw the laser off the track entirely, the transport’s own tracking servo can lose the track outright — holding the last valid samples in its buffer and repeating them, or jumping the pickup ahead to reacquire the track. That is the stutter and the skip a scratched disc is remembered for — a mechanical failure, not an algorithmic one: correction and concealment are choices about damaged data; a lost track is the pickup losing the disc, not the code losing it.
None of this was optional. A pressed compact disc’s ordinary handling — fingerprints, dust, the fine scratches of years in a sleeve — guarantees bit errors on nearly every play; a format with no correction scheme would have been unlistenable almost immediately. CIRC’s job was to make an imperfect physical medium behave like a reliable digital one, and its layered response — correct, then interpolate, then mute, then let the mechanism do what it still can — is the honest hierarchy of a system built to fail gracefully rather than never fail at all.
Yumeo’s CD-9 is a compact disc player from exactly the moment that hierarchy becomes audible — a disc worn enough that correction alone is no longer doing all the work. Orrery’s restoration is modeled from a specific surviving unit:
Yumeo CD-9 · compact disc player · restored from unit no. C-3090622 · Yumeo, Japan, 1988
A failing disc, held at the moment it gives up — the interpolation, the brief digital silences, the stutter. The concealment logic is the real subject; a random glitcher only imitates it.
The instrument: https://orrery.dk/yumeo/cd-9/
-
2026-08-30
What ATRAC compression is
ATRAC — Adaptive TRansform Acoustic Coding — is the compression scheme Sony built for MiniDisc, launched in 1992 to fit a little over an hour of digital audio on a disc a fraction of a compact disc’s physical size. ATRAC is lossy: it does not store the recording, it stores a version edited down by a model of what the human ear can and cannot hear, at roughly a fifth of the original data. That is the reason a MiniDisc recording, even at its best setting, has a characteristic softness — a slight rounding of transients, a stereo image that can read a touch narrower than the source — audible mostly in what stops being there rather than in anything added.
At its standard rate, ATRAC compresses 16-bit, 44.1-kilohertz stereo audio — about 1.4 megabits a second, uncompressed — down to roughly 292 kilobits a second, while remaining decodable in real time on a battery-powered deck built in the early 1990s. Near-CD duration and playback quality, at about a fifth of the data a CD spends.
The coder splits the audio spectrum into three frequency bands with a filter bank — roughly 0–5.5 kHz, 5.5–11 kHz, and 11–22 kHz — then applies a modified discrete cosine transform (MDCT) inside each band, turning short blocks of the waveform into frequency coefficients. The transform’s window length adapts to the material — long windows for steady, tonal passages, where fine frequency resolution matters; short windows around sharp attacks, where fine time resolution matters more, to limit the transform’s tendency to smear a hard transient across the whole block.
A psychoacoustic model then decides how many bits each block floating unit — a small group of adjacent frequency coefficients, not each coefficient singly — is worth. Loud content masks quieter content nearby in frequency and immediately after it in time — the ear cannot separate them — and the model spends its limited bit budget on what it predicts will remain audible, quantizing or discarding the rest. At ATRAC’s standard rate the model is usually right: the loss sits in cymbal decay, room tone, the trailing edge of a transient, the outer edge of the stereo field, rather than in anything a casual listener would flag as missing.
At the long-play modes added in 2000 — LP2 at 132 kilobits a second, LP4 at 66 — a later coder, ATRAC3, has to fit twice or four times as much material into the same space. At LP4 it runs out of room and codes the two channels jointly, folding the stereo image toward the centre: spatial information traded for bit budget once there is nothing else left to trade.
MiniDisc’s whole premise depended on this trade: an album on a disc a fraction of a compact disc’s physical size, played back inside a portable deck’s power and processing budget, using early-1990s digital signal processing. Uncompressed 16-bit audio was never going to fit at that size; a perceptual coder was the only way to make a recordable, pocketable disc format viable at all. ATRAC’s job was never bit-exact reproduction. It was fitting the perceptually important part of a recording into the rate the format could actually carry.
Yumeo built its MD-1 for exactly this generation of format — an ATRAC recorder sold as a nearly perfect digital medium at a moment when digital itself still felt like the promise. Orrery’s restoration is modeled from a specific surviving deck:
Yumeo MD-1 · ATRAC recorder · restored from unit no. M-2170534 · Yumeo, Japan, 1994
The restored unit — number 2,170,534 off the line — was, like most surviving MD decks, barely used: obsolete before it could wear out. The character is the codec, and we kept all of it.
The instrument: https://orrery.dk/yumeo/md-1/
-
2026-08-28
On the bench: 4400 · Quantization Analyzer.
Tekura Type 4400 · Quantization Analyzer · restored from Ser. No. 00417-A · Tekura, Oregon, USA, 1988
Eight bits, truncated — The phrase runs dry for five seconds, then the bench default: Bits at 8, Quantize set to Truncate, Dither off. The grain sits in the decays, where the level falls and the staircase shows.
The instrument: https://orrery.dk/tekura/4400/
-
2026-08-23
The engine we could not borrow
The Yumeo MP-3 is a 1999 perceptual coder, and restoring it honestly meant encoding MPEG-1 Layer III — not decoding it, encoding it, with the era’s arithmetic intact. The swirl at low bitrate is not a defect to be cleaned away; it is the object under restoration.
An encoder is a tool like any other on the bench, and the bench prefers to borrow before it builds. But every Layer III encoder we could reach carried a licence that could not travel inside a sealed instrument. No fault of theirs — those licences protect a different kind of work. Ours simply could not use them.
So the bench built its own. First a reference in plain arithmetic, slow and readable, every table transcribed from the standard and checked twice. Then the working engine, held against that reference until the two produced the same bytes — not similar bytes; the same bytes. The reference stays in the case beside the tool, the way a conservator keeps the measurement beside the repair.
We have named it epicycle, for the small circle in the old orbital models — an elegant approximation the world outgrew, which still predicts beautifully. That seemed right for MP3.
The engine now drives the restoration at the bench, and it will carry the MP-3’s next issue. And because the licence problem it solved is not ours alone, the source is public: MIT-licensed, tests and reference included, at github.com/orreryarchive/epicycle. Take it, verify it, use it.
Note: Orrery is a work of fiction — the houses, the catalogue’s histories, and Iver Sandholm are invented. The encoder is not. It was written at this bench with AI assistance under human direction, and every claim in its repository can be re-proven from source. The instruments it serves are real software, really sold and supported. Only the story is a story — and we think the work is better for being checkable where it matters.
-
2026-08-23
Why GSM calls sound underwater
A 1990s mobile call does not sound bad at random. It sounds bad in a specific, repeatable way: a faint gurgle under vowels, sibilants that thin into a soft hiss, a texture closer to a voice heard through a wall of water than through wire. That quality has a name and a cause. It is RPE-LTP — Regular Pulse Excitation, Long-Term Prediction — the speech codec GSM adopted to carry a call at 13 kilobits a second, and the underwater grain is not a defect in some particular handset. It is the sound of the compression working as specified.
GSM’s full-rate codec, standardized as GSM 06.10 by the European GSM standardization group (later folded into ETSI), takes voice sampled 8,000 times a second and fits each 20-millisecond frame — 160 samples — into 260 bits. That is 13,000 bits a second — a fraction of what a landline call carries, let alone a CD. Nearly everything about the underwater sound follows from how that small budget gets spent.
RPE-LTP does not transmit the waveform. It predicts it. A short-term linear predictor models the shape of the vocal tract for that frame — the resonant shape of mouth and throat that gives a vowel its character. A second, long-term predictor models pitch: the repeating pulse of the vocal cords in voiced speech, spaced by the pitch period. Together the two predictors reconstruct most of what a vowel sounds like from very little data, because a vowel is repetitive and a handful of numbers describes it well.
What is left after both predictions — the residual, the part neither predictor accounted for — is where an unpredictable sound actually lives: consonants, sibilants, breath, background noise. RPE-LTP does not send that residual whole. It sends a regularly spaced subset of it — every third sample, thirteen out of forty, on one of four candidate grids chosen for whichever carries the most energy. The decoder puts those thirteen pulses back on the grid and zeros between them, and leaves the synthesis filters to make a signal out of that. For a clean vowel the approximation holds and the loss is barely audible. For a fricative or a hard consonant, where nearly all the information sits in that discarded residual, the approximation runs out — and what returns is a wavering, band-limited, faintly liquid texture. It is the codec’s honest best guess at a signal type its architecture was never built to carry well.
Thirteen kilobits a second was not a stingy choice; it was the ceiling. Channel coding turned it into 22.8 kilobits on the air, and eight of those shared a single 200-kilohertz carrier — decoded in real time on mobile silicon with a fraction of a modern phone’s processing budget. RPE-LTP was the trade that fit inside that ceiling: spend the bits on what speech does predictably — pitch, vocal-tract shape — and let everything else arrive as a coarser approximation. The alternative was not a clearer call. The alternative was a network that could not carry the call at all.
Salo Oy built the network-side test hardware that lived with exactly this trade in the mid-1990s — codec boards and comfort-noise generators that told a Nordic carrier’s engineers what a call would sound like before it reached a customer’s ear. Orrery’s restoration is modeled from one of those units:
Salo 3310 · Full-Rate voice codec · restored from unit no. CU-9426-0173 · Salo Oy, Salo, Finland, 1994
The watery quality is the RPE-LTP algorithm working exactly as designed. We have not clarified it; clarity was never the assignment.
The instrument: https://orrery.dk/salo/3310/
-
2026-08-18
On the bench: 3310 · Full-Rate GSM Codec.
Salo 3310 · Full-Rate voice codec · restored from unit no. CU-9426-0173 · Salo Oy, Salo, Finland, 1994
The call, filtered — The note is read dry for six seconds; the full-rate codec then takes the line. What arrives is the 1994 call — half-swallowed, faintly underwater, exactly as designed.
The instrument: https://orrery.dk/salo/3310/
-
2026-08-15
The record opens.
Orrery keeps a public record now. Entries are short and factual: what arrived at the bench, what returned to service, what changed in the catalogue. The bulletin remains the way to be written to; the record is the way to look in.