From 36c737f45efc4f9459a7923d62ccf2cf5bf2686a Mon Sep 17 00:00:00 2001 From: Craig Jennings Date: Thu, 6 Aug 2026 11:50:47 -0500 Subject: docs: record how the ambience audio was built - Keeps the stream-copy loop technique and its measurements. - Notes the amix normalize=0 trap that silently discards the mix ratio. - Records the loop-period arithmetic behind the crossfade. - Explains why the openly-licensed set sounds better but its rain is wrong. The audio itself stays untracked. Claude-Session: https://claude.ai/code/session_01WpQhUr9mNrXss3HiaN6G2s --- docs/design/2026-08-03-ambience-audio-handoff.org | 79 +++++++++++++++++++++++ 1 file changed, 79 insertions(+) create mode 100644 docs/design/2026-08-03-ambience-audio-handoff.org (limited to 'docs') diff --git a/docs/design/2026-08-03-ambience-audio-handoff.org b/docs/design/2026-08-03-ambience-audio-handoff.org new file mode 100644 index 00000000..045e44ad --- /dev/null +++ b/docs/design/2026-08-03-ambience-audio-handoff.org @@ -0,0 +1,79 @@ +#+TITLE: Ambience audio — build notes and source findings +#+AUTHOR: Craig Jennings +#+DATE: 2026-08-03 + +Build notes for the ambient audio the music config plays. The audio itself +is not tracked here; it lives at =~/documents/sounds/ambience/= and the +playlist reaches it by relative path. This file is the technique and the +measurements, which are the parts worth keeping. + +* What these are + +Three 1-hour ambient tracks built 2026-08-03: + + cafe-1h 29 MB cafe murmur only + rain-1h 25 MB rain only + rain-plus-cafe65-1h 43 MB both, cafe at 65% of the rain + +* How they were built + +Sources are two short loops — cafe at 1:50 and rain at 2:32, both 24 kHz +stereo at 56-64 kbps. + +The two pure tracks are =ffmpeg -stream_loop= with =-c copy=, so they carry +no second-generation encoding loss: bit-identical to the source repeated +about thirty times. Only the mix is re-encoded (libmp3lame 96k, 24 kHz to +match source). + +The mix is rain at unity and cafe at =volume=0.65=, combined with +=amix=normalize=0=. + +*The =normalize=0= is the part to remember.* =amix= normalizes by default, +which halves both inputs and throws the 65% ratio away — the thing you were +setting is silently discarded and the output still sounds plausible. + +Verified after building: peak -1.1 dBTP (no clipping), mean -23.3 dB. + +=cafe-1h= is 3599.26s rather than 3600 because a stream copy cuts on an MP3 +frame boundary. I took the lossless copy over the last 0.74s. + +* A better-sounding source set exists + +There is also a set built from openly-licensed field recordings, at +=~/documents/sounds/ambience/cc0-*.mp3= — rain from "Lluvia En Riumors" +(CC0 1.0, 18:08, lossless FLAC source) and cafe from "Cafe, Aldeburgh, UK" +via radio aporee (Public Domain Mark 1.0, 17:33, 320 kbps). + +They sound better than the short loops for a structural reason: 44.1 kHz +throughout, and an hour needs about four passes rather than thirty, so there +is far less audible repetition. + +** Open thread — that set's rain is too percussive + +It was recorded on hard pavement, so you hear drop impacts you would never +hear from inside a cafe. + +Measured: its energy above 2 kHz sits only 0.4 dB below full-band, against +-19.9 dB for a real rain-through-a-window recording. + +Filtering the pavement rain to match does not work — it strips to sparse +thumps, because that recording's rain character lives almost entirely above +1 kHz. The fix is a different source recorded indoors, which is unresolved. + +Relevant if a rain-only mode ever gets built. + +* Note on the loop periods + +Measured separately when the 15-minute loops were cut: the periods are exact +rather than estimated, because the files are literal stream-copy repeats, so +the period is found by locating where the byte stream repeats. cafe is +109.680s (4570 frames), rain is 151.584s (6316 frames). + +gcd(4570, 6316) = 2 frames, so the two layers coincide only every 14,432,060 +frames — 96.2 hours. No practical length wraps both cleanly, which is why the +15-minute mix uses a 3s equal-power (qsin) crossfade folding its tail over its +head so F(0) = F(L), rather than a naive cut. + +Measured across the wrap against the typical adjacent sample step inside the +file: a naive 900s cut gives 4875 against 248 (19.7x, an audible click); the +crossfaded version gives 236 against 229 (1.03x). -- cgit v1.2.3