The EMU3 sample record's pointer block declares a channel count, and roughly 2 700 of the 14 738 E-mu samples on the reference shelf declare two channels. The payload is a block split — all of the left channel, then all of the right — so reading it as one mono stream concatenates the channels into a file twice as long as the sound.
Found while decoding the loop points in #31 (ADR-0025). Not fixed there: correcting it moves audio, and D17 was scoped to add metadata without touching a byte of payload.
Why this was missed
ADR-0020 rejected the channel-count reading on a measurement that tested the wrong hypothesis: it de-interleaved payloads as LRLR and found the roughness roughly doubled. That is a sound refutation of interleaved stereo and says nothing about a block split, which read as mono is one continuous waveform with a single join in the middle. The rejection is annotated in place; the rest of ADR-0020 stands.
The evidence
Median RMS-envelope correlation between the two halves of a payload, split by what the record's own pointers declare:
|
esi32-gm |
eiiix-1 |
eiiix-2 |
protozoa |
analogia |
studio |
vitous |
| two channels declared |
0.99 |
0.71 |
0.95 |
0.55 |
0.91 |
0.96 |
0.82 |
| one channel declared |
0.13 |
0.16 |
0.59 |
0.26 |
0.21 |
0.22 |
— |
Envelope rather than sample-level correlation: the two channels of a real recording differ in phase and agree in shape.
Records declaring two channels, per disc: esi32-gm 28, eiiix-1 641, eiiix-2 598, protozoa 27, eiv-analogia 279, eiv-studio 320, eiv-vitous 828 (all of it).
What a fix involves
start_R == start_L + P/2 is the test. The two halves interleave into a stereo WAV, which wav.py already writes and stereo.py already interleaves for the name-paired case. Watch for:
- Every E-mu sample count and payload digest moves for those records.
tests/test_discs.py pins both per disc; they will need re-measuring, and the reason recorded.
- The loop points do not move.
(pointer − start) / 2 is a per-channel frame index either way.
eiiix-2's one-channel figure of 0.59 is the weakest separation in the table and should be understood before the gate is trusted.- The interaction with ADR-0017 name-pairing, which is a separate mechanism these discs also use — both are real, and that is part of why one hid the other.
Written up in docs/formats/emu3.md under "Stereo: the payload is split, not interleaved".
The
EMU3sample record's pointer block declares a channel count, and roughly 2 700 of the 14 738 E-mu samples on the reference shelf declare two channels. The payload is a block split — all of the left channel, then all of the right — so reading it as one mono stream concatenates the channels into a file twice as long as the sound.Found while decoding the loop points in #31 (ADR-0025). Not fixed there: correcting it moves audio, and D17 was scoped to add metadata without touching a byte of payload.
Why this was missed
ADR-0020 rejected the channel-count reading on a measurement that tested the wrong hypothesis: it de-interleaved payloads as
LRLRand found the roughness roughly doubled. That is a sound refutation of interleaved stereo and says nothing about a block split, which read as mono is one continuous waveform with a single join in the middle. The rejection is annotated in place; the rest of ADR-0020 stands.The evidence
Median RMS-envelope correlation between the two halves of a payload, split by what the record's own pointers declare:
esi32-gmeiiix-1eiiix-2protozoaanalogiastudiovitousEnvelope rather than sample-level correlation: the two channels of a real recording differ in phase and agree in shape.
Records declaring two channels, per disc:
esi32-gm28,eiiix-1641,eiiix-2598,protozoa27,eiv-analogia279,eiv-studio320,eiv-vitous828 (all of it).What a fix involves
start_R == start_L + P/2is the test. The two halves interleave into a stereo WAV, whichwav.pyalready writes andstereo.pyalready interleaves for the name-paired case. Watch for:tests/test_discs.pypins both per disc; they will need re-measuring, and the reason recorded.(pointer − start) / 2is a per-channel frame index either way.eiiix-2's one-channel figure of 0.59 is the weakest separation in the table and should be understood before the gate is trusted.Written up in docs/formats/emu3.md under "Stereo: the payload is split, not interleaved".