819a2757bc5ecd85cbf0d673abdd5b87a42f4b4d
5
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 KeelrLeek
|
52f5878297 |
loudness: a first play starts three quarters of a decibel under the prefix
The prefix estimator is BIASED, not merely noisy, and the sign is known: a quiet opening under-reads the track, so the head asks for too much gain far more often than too little. Subtracting a constant moves the whole distribution toward the truth, so a first play begins nearer its real level and the correction that follows has less ground to cover. Measured by a new `head_start_offset_sweep` over 1046 tracks sampled across 182 artists — at most six per artist, spread across each discography rather than taken from the front, since openers are where intros are least typical and intros are the whole source of the bias. The cached corpus it replaces was a dozen artists and put the median at +3.48 dB. Sampled diversely it is +1.55, so the old number was overstating the bias by more than double. The size comes from the MARGINAL trade, not the smallest mean error: step loud fixed quiet created ratio 0.0 -> 0.5 -89 +41 2.2 : 1 0.5 -> 1.0 -83 +85 1.0 : 1 1.0 -> 1.5 -61 +132 0.5 : 1 Only the first step is clearly worth taking. mean|err| bottoms out at 1.5 dB instead and is the wrong objective: it weights an error that begins decaying at once the same as one that waits for Analysis::Complete, which on the streamed first plays this path exists for is download-bound and therefore late. A constant is deliberately a weak lever. The per-artist median error runs from -4.58 dB (Boards Of Canada) to +7.03 dB (Ponytail); no single number fits everyone, and an offset shifts that spread without narrowing it. The levers that would are letting the follow recover UPWARD, and predicting the offset per track from the prefix. Both still open. The overlay reports it as a MOVE — what the prefix asked for, where the track actually started, and a ghost tick on the dB scale at the prefix position, so the gap to the applied gain IS the head start. Gated on `analyzing`: at Analysis::Complete the player releases to the whole-track gain while loudness.gainDb goes on reporting the prefix value, which it keeps for the next play. Without that condition the line advertised a correction that had already been released. |
||
|
KeelrLeek
|
1e35c5e039 |
loudness: correct the gain during the track instead of only guessing at its start
A streamed first play is gained from a prefix; every later play is gained from the whole track. Measured across the 89 cached tracks, a 10s head asks for a median 3.4 dB too much gain and is more than 1 dB hot on 68 of them — that gap is the "first track is so much louder" complaint. The curve never flattens either: even a 60s head is still ~1 dB out with a 6 dB worst case, so no window size fixes it. Live traces then showed something worse. The hard deadline was anchored to the play REQUEST, so on a peer that took a moment to start sending, the 2.5s ran out before a single gating block existed and the track began at unity — no normalization at all, three plays out of four on one run. Releasing early bought nothing in exchange: playback is blocked on the same bytes, so the gate was discarding the measurement without making sound arrive sooner. - The deadline now starts at the first byte, with an absolute 4s ceiling from the request so a slow peer cannot hold audio that is ready to play. - The audio thread's slew drops 2.5 -> 0.25 dB/s. Mid-track correction was switched off in July because the glide was audible; at a tenth the rate it is not, and being 3 dB wrong for a whole track is the larger error. - The correction follows the running measurement from its first publish and may only ratchet DOWN until the track completes. A convergence test lived here and was deleted: an integrated average over 30s cannot move much when one more second is added, so "3 publishes within 0.3 dB" was measuring the inertia of a long average, fired at its floor every time, and declared settled while 1.5 and 2.1 dB from truth. The protection was never the gate — it is the slew rate and the ratchet. - Seeking lands the pending correction immediately. The slew is slow only because a gain change is audible against continuous music, and a seek is already a discontinuity. Observability: SOULSTREAM_LOUDNESS_TRACE=1 reports every commit with the seconds analyzed, which limit fired, the hold, and the bytes received; the debug overlay gains a SLEW line showing applied gain, target, distance and ETA, because one number cannot distinguish "settled" from "in flight". Also adds an #[ignore] corpus harness that sweeps head-window bias in 30s. |
||
|
KeelrLeek
|
fa3cbc64ca |
loudness: a prefix may lower a track's volume but never raise it
A track that opens on a lone guitar and swells into a full chorus played that
chorus far too loud. The head of an unmeasured stream was being treated as a
measurement of the whole song: at ten decoded seconds the analyzer's running
integrated value became the gain for the entire play, and a quiet opening
under-reads, so the estimator asked to turn the track UP and the song then
arrived at its real level already amplified. `PRELIM_MAX_BOOST_DB = 3.0`
bounded that mistake without preventing it, and the three decibels land
exactly on the loudest part of the song.
Measured on the 69 complete tracks in the stream cache, with the app's own
meter and its own gain policy, comparing what the 10 s head asks for against
what the finished track asks for:
prefix error in applied gain (+ = plays too loud)
1s min -4.6 p50 +7.7 p95 +15.8 max +17.0 too loud: 61/69
3s min -0.9 p50 +5.2 p95 +15.7 max +17.0 too loud: 57/69
10s min -1.5 p50 +2.9 p95 +14.6 max +17.0 too loud: 57/69
The head is wrong in one direction and wrong by a lot. So it may only ever
cut: `PRELIM_MAX_BOOST_DB` is 0.0, and a boost has to be earned by a
measurement that reached the end of the media.
**Cuts-only rather than the flat unity the design called for.** The design
doc's stated first-release policy is that an unknown partial stream stays at
0 dB, with cuts-only left as a later experiment. Measured against this
corpus, unity is the worse of the two in the direction being complained
about — first-play loudness relative to the −14 LUFS target:
unity (doc's policy) p50 +5.6 p95 +8.5 LOUDEST +10.6 >3dB over: 57/69
cuts-only p50 +2.7 p95 +6.7 LOUDEST +8.0 >3dB over: 31/69
old policy, capped +3 p50 +2.7 p95 +9.5 LOUDEST +11.0 >3dB over: 32/69
verified (later plays) p50 +0.0 p95 +0.0 LOUDEST +0.0 >3dB over: 0/69
Unity would hand back the cut that loud-opening tracks already get today:
"Spread Eagle Cross the Block" reads −3.4 LUFS over its head and −3.4 over
the whole track, and is cut 10.6 dB — under unity its first play would be
10.6 dB louder than it is now, which is the reported bug wearing a different
hat. The tail risk that argues for unity is a loud opening over a quiet body,
where the head over-cuts; on this corpus that case tops out at 1.5 dB (My
Bloody Valentine, "Only Shallow": head −11.3, track −12.8) and the existing
−15 dB floor bounds it. The corpus is one library and contains no fortissimo
opening over a pianissimo body, so it bounds the error it can see, not the
error that exists. Revisit if a cut ever sounds wrong.
**Only a whole programme may be cached.** `analyze_loudness` returned the
same `Option<(f64, f32)>` whether it reached the end of the media or the
bytes simply stopped, and `on_track_started` stored it either way — with a
comment saying it was valid "whatever happened to this one". It is not:
`GrowingFile::read` maps an aborted transfer to `Ok(0)`, a graceful EOF, so
skipping a song mid-download recorded whatever its intro measured, filed
under the full declared remote size, and every later play boosted the whole
track off it by up to `MAX_GAIN_DB` — the same bug as above, but permanent
and three times larger. It now returns `Analysis::Complete | Incomplete`,
where completeness is decided by asking the transfer how it ended
(`StreamProgress::terminal`) rather than by asking a reader that cannot tell,
and only `Complete` reaches the cache, the preload boundary gain, or a
preview clip's gain.
That path was reachable but had not fired. Re-measuring every cached track
against its stored value found no disagreement, and of the ten interrupted
transfers still sitting in the stream cache, none had left a loudness entry
at all — the analyzer normally wins its race with the download and either
finishes properly or is cancelled before the abort reaches it. So the
existing cache is not quarantined and not wiped: of 141 entries only 6 ask
for any boost above 0.5 dB, four of those are present on disk and verified
correct by a fresh scan, and the other two ask +1.4 and +1.0 dB. Discarding
141 good measurements to protect against a hazard that measurement says is
not there would cost more than it saves, and any record that is wrong is
overwritten by the next complete play.
**The audible start is now bounded from the play request.** Both existing
windows are measured from the first measurement, and a track opening below
the −70 LUFS gate never produces one, so silence or a long fade-in held
playback until the analyzer found audio loud enough to measure — on a slow
stream, a wait with no upper bound. `PRELIM_HARD_DEADLINE` (2.5 s) releases
the start regardless, at the best cut the evidence so far justifies. This
costs nothing: the player blocks on the growing file anyway, so releasing
early hands it the start rather than making it wait. Verified against a
223 MB FLAC whose first bytes had not arrived at 2.2 s.
**Stop now cancels the measurement window.** Found while verifying the
deadline: `player_control("stop")` never touched the loudness generation, so
a track stopped during its window could be handed its deferred start a moment
later and begin playing after the user had stopped it. Pre-existing, but the
deadline makes it reliably reachable, so `on_playback_stopped` orphans the
analyzer and the held start.
Not done, and deliberately: true peak and the −1.0 dBTP ceiling (the meter
measures sample peak, so `PEAK_CEILING` is still 0.98 sample), the versioned
measurement cache with real asset identity, ReplayGain/R128 metadata, the
output limiter, track-bound gapless gain, and queue-ahead analysis. Those are
later phases and none of them is needed to stop the over-boost.
Verified live, driving the running app:
quiet intro, first play head −27.4 LUFS → 0.0 dB (was +3.0 dB)
quiet intro, second play track −6.0 LUFS → −8.01 dB from the cache
loud intro, first play head −3.4 LUFS → −10.64 dB cut still applied
Soulseek stream, cold head −18.8 LUFS → 0.0 dB, track measured −9.0
and cached for the next play without touching
the gain in force
stop mid-window no start behind the stop, no cache entry written
The debug overlay now names what the gain was decided from — "whole-track
measurement (cache)", "prefix (cut only)", "unity (nothing measured yet)" —
because the difference is invisible in the numbers alone, which is how this
went unnoticed. Settings no longer claims "boosts never clip", which nothing
in the output path enforces; it describes the actual guarantee instead. The
module header's claims that the gain glides mid-track and that previews are
not normalized were both false and are gone.
150 Rust tests and 85 JS tests pass.
|
||
|
KeelrLeek
|
28aba3a3b3 |
feat: normalize preview clips to -14 LUFS to match the library
Deezer preview clips are cut from the raw master (not pre-normalized), so they played several dB hotter than the user's normalized -14 LUFS files — measured across one artist the previews ranged -0.1 to -8.8 dB from target, an 8.7 dB spread, so a flat offset would have been wrong either way. Each clip is now measured with the same BS.1770 meter the main player uses (preview.rs measure_clip_gain -> loudness::normalization_gain_db) right after it is written in preview_resolve -- off the async worker, and fully off the playback path on a prefetched neighbour. The linear gain is stored in the lease and the reuse cache, and rides UNDER the app-volume mirror per voice (effective = target_vol * clip_gain), so it composes with the slider, mute and the crossfade (each clip crosses in at its own matched level). Gated on the volume_normalization setting; reset on dispose. Debug overlay PREVIEW row now shows the active clip's norm dB. Verified live: per-clip gains -0.1..-8.8 dB applied; mute still zeroes the preview; 39 app + 12 player tests pass. |
||
|
KeelrLeek
|
793f78370e |
feat: crossfade preview suite, loudness normalization, debug overlay
Preview modal audio, five rounds: - True track-change crossfade: scheduleSettle no longer silences the outgoing clip at t=0; preview_load's equal-power fade owns both ramps, gated on the incoming voice's first Playing event (audible_marks), with failure/overstay fallbacks. CROSSFADE_MS = 340 — locked to the reel's card-slide so audio and visuals land together. - Two-neighbour lease prefetch + a 15-min on-disk clip byte-cache (recent_clips): steps in BOTH directions crossfade in ~15 ms; revisits never re-download. - Adaptive settle (isolated steps fire immediately, bursts coalesce) and anticipatory auto-advance (step at clip end minus 400 ms) so auto-play crossfades over the tail instead of gapping. - Session resume memory: revisited tracks continue where they left off (decoder-level start offset via play_local_from/resume_at); finished tracks restart from the top. - Open/close crossfades with the MAIN player: fade-pause with the desired-paused latch landing at ramp end, fade-resume from the paused position; MAIN_FADE_GEN guards; focus handshake relaxation documented in focusBridge. Volume normalization (Settings > Playback, default on): - Dependency-free BS.1770/EBU R128 meter in the player crate (K-weighting, 400 ms gated blocks, incremental) + analyzer that reads the same growing file playback streams from — first plays are measured at download speed. - Gain layered under user volume; audible start GATED on a preliminary measurement (10 s of audio or 800 ms budget) so the level is right from the first sample and never changes mid-track; prelim boosts capped 3 dB, peak-aware clipping guard, persistent loudness-cache.json. - Gapless boundaries get their own sample-accurate gain via SetNextNormGainDb, fed by preload analysis. Debug overlay (Ctrl+Shift+D): process-tree CPU / PDH GPU / working-set + private-commit RAM / network rates / FPS / current-track download progress read from the backend's own StreamProgress / live normalization state / preview session phase. Polls only while visible. 131 JS + player + app tests green; every path verified live over CDP. |