Sylpheed/docs/re/structures/intro-audio-output-census.md

What the game actually emits over the boot intro — five live channels, not a stereo mix

Classification: measured. Xenia Canary, 2026-08-30, one boot, no input. Groundwork for the port's ask #4 — it does not settle #4, and the part it does not settle is named at the bottom.

Why this capture exists

The port ships one of ADV's three concurrent voice streams and its authored/audio.json records that as known wrong, held deliberately (voice-three-streams-are-concurrent.md). Neither "take one" nor "sum them" is established. The only thing that can settle it is what the game emits, so: record it.

The capture

Recipe followed exactly from audio-capture-alsa-file-tee.md — ALSA file tee in front of a paced pulse slave, --gpu=null, both mutes off:

run-canary --apu=alsa --mute=false --gpu=null --xma_param_probe=true

148.02 s, 6 channels, float32, 48 kHz.

✅ Provenance is the XMA probe, not a screenshot — which is the right evidence for an audio question, because it witnesses the thing being recorded. The log carries ADV's three contexts byte-exact against the disc:

ctx	packets	byte_size
0	632	1 294 336
1	546	1 118 208
2	572	1 171 456

then two more — 1 150 976 / 1 269 760 — which are the documented BGM_102 pair. So the movie's voice decoded exactly as in the runs this corpus already records.

✅ Capture quality: 0.15–0.16 % silence on the five live channels, against the 0.31 % the recipe page records for its own clean run. Cleaner than the reference.

What is in the six channels

⚠️ ALSA's channel order, not WAV's: captured i holds source [0,1,4,5,2,3], so the labels below are FL FR BL BR FC LFE. Deterministic and invertible.

ch		peak dBFS	rms dBFS	% silent
0	FL	−3.65	−22.26	0.16
1	FR	−2.22	−20.89	0.15
2	BL	−4.41	−24.77	0.15
3	BR	−11.65	−36.06	82.18
4	FC	−4.57	−24.81	0.16
5	LFE	−4.66	−24.38	0.16

Five channels carry real, distinct content; one (BR) is 82 % silent and 11–15 dB down. No channel is a copy of another — the largest pairwise correlation is 0.70 (FL/FR, which is what a stereo pair looks like), then 0.52 and 0.56.

🔴 What this already rules out

• Not a stereo mix. Five independent channels are being emitted.
• So "ship one stream" cannot be right, which the corpus already suspected but had not observed from the output side. The port's held-wrong value stays wrong.

⚠️ It does not follow that summing is right. This says the output is multichannel; it says nothing yet about which disc stream lands where.

⚠️ And do not read the 6 channels as proof the game is 5.1. CONTAINER-NOTES records that AudioDriver::kFrameChannelsDefault = 6 is a hardcoded Xenia constant — the format is Xenia's, only the content is the guest's. What is evidence here is that five of those six differ from each other, which a stereo guest cannot produce.

🟡 What is NOT settled — the stream → channel mapping

The decisive step is a cross-correlation of each captured channel against each of ADV's three decoded streams. That has not been run. Until it is:

• which stream feeds which channel pair is unknown;
• whether FC carries the dialogue is consistent with the corpus's centre-channel finding but not re-established here;
• why BR is near-silent is unknown — a genuinely unused channel, a stream that ends early, or a decode that failed are all open.

⚠️ One boot, one movie. And --gpu=null means no video, so nothing here is cross-checked against what was on screen.

The raw capture is 170 MB and is not committed; it went over share to the port as 1788077587-9f2e30af1c98-capture.raw. Per-channel numbers: ../data/intro-audio-channel-census.txt.