This page is Brandon's reading map for his music-audio work. He splits songs into their separate instrument tracks and makes voice readings loud enough to hear. This page lists the books that would teach the hard parts and the tools he already built that do the job. Tap Grown-up mode to see the full list.
Books worth reading and skills already in the library, mapped to the exact decisions in the stem-decomposition and Etsy-loudness work.
Fit scores below are relevance to these problems, not the book's general fame.
Raw export measures -33.86 LUFS at 130 kbps. Target is -16. Customers say they cannot hear the reading.
Demucs then DrumSep produce lanes. A reconstruction near 1.0 proves the pieces add back, not that each holds the right content.
GCC-PHAT, polarity, and offset. The -44.9 / -8.4 / +4 ms bass conflict is a periodicity question, not a by-ear one.
40 to 160 Hz summation, mono compatibility, kick-fundamental retention, and whether a near-empty lane belongs in the mix.
Twelve books, three per problem, ranked. The top pick in each group is the one that speaks directly to the decision you are stuck on.
The -33.86 to -16 LUFS gap, and the perception behind a reading that reads as inaudible.
The loudness reference. Defines LUFS, the K-system, EBU R128, true-peak, and how to hit a delivery target on purpose instead of by luck. It is the exact vocabulary of your -33.86 to -16 problem.
Touches: fix_audio_levels.sh target, and why GarageBand exports land ~18 dB quiet.Loudness perception, masking, and threshold. Explains why a quiet reading sitting close to the noise floor is heard as "cannot hear it" on a phone speaker even after a small boost.
Touches: whether -16 is loud enough for real playback conditions.Practical loudness, dither, and delivery specs for consumer playback. The hands-on companion to Katz for setting your final m4a target.
Touches: the delivered-file spec buyers actually open.Source separation quality, and the metrics that tell content apart from a well-formed empty lane.
The reference text on the field Demucs and DrumSep come from. BSSEval (SDR, SIR, SAR), SI-SDR, mixture consistency, and the evaluation traps. This is the book behind your point that recon near 1.0 is a sanity check, not proof of content.
Touches: the museval SI-SDR / SIR / SAR check queued on the drum lanes.Onset detection, chroma, tempo, and DTW alignment, all in the MIR math your key-vote and window-pick steps already use. The cleanest single source for the analysis layer of the pipeline.
Touches: the pYIN note-continuity check on the bass lane.The perceptual theory of how a listener pulls one source out of a mixture. Grounds what "clean" means to an ear and why residual bleed matters even when a number looks fine.
Touches: your mandatory audible-review gate.Cross-correlation, polarity, offset, and the periodicity that makes a bass-only offset ambiguous.
Correlation, cross-correlation, and phase, with the periodicity math behind "smaller bass-only offsets are cycle aliases of a periodic low end". The book that settles why you cannot pick between offsets a bass cycle apart by ear.
Touches: the -44.9 vs -8.4 vs +4 ms bass-offset conflict.Plain-language correlation, convolution, and FFT. The fastest on-ramp to the transient-anchored narrow search you already built, and it costs nothing.
Touches: the one-coarse-offset-then-narrow-search model.DSP taught in Python, matched to your scipy and librosa toolchain. The best fit for coding the composite-scoring test as an actual script.
Touches: the kick+bass summation scorer.Low-end summation, mono compatibility, and the mix-check that resolves an offset better than a correlation number.
Phase alignment, low-end summation, mono compatibility, and mic bleed, written for exactly the gear you have. This is the composite-aware test the brief prescribes for the bass resolver, in book form.
Touches: build kick+bass at each offset, score 40 to 160 Hz summation and mono compat, take the fullest low end.Balance, frequency-space arrangement, and practical bleed handling. The reference for whether a near-empty lane earns a place in the composite.
Touches: the guitar 7.31 vs 8.07 keep-or-drop decision.The DSP of every process you run on a lane after alignment: filtering, dynamics, and the band-limited energy work your tie-break needs.
Touches: loudness-matched, band-limited comparison of the two guitar candidates.Grouped by what they do for this work. Each carries its invoke command. These are the current library, not a wish list.
The direct match. Pull and verify a source master, drum-aware song mapping of a continuous jam, serial Demucs 6-stem then DrumSep MDX23C 5-piece, frame-matched packages, SHA256 provenance, atomic per-stage status. This is the harness your 8.07 / 8.00 / 7.31 run lives in.
/jam-stem-rebuildReads a song supplied as five stems through a songwriting lens after the separation step. The interpretation layer above the decomposition.
/five-stem-song-analystPer-stem inspection of content and quality. Use it on each lane before trusting a package.
/stem-analyzerFrequency content per stem. This is your read on the 40 to 160 Hz low end and where the bass energy actually sits.
/stem-frequency-mapAudition stems. The audible review the brief calls mandatory runs through here.
/stem-playerDemucs plus per-stem transcription (crepe, basic-pitch, onset). Gives ground-truth MIDI to check a lane against.
/audio-stems-to-midiMix-decision reference. The in-library companion to Senior and Owsinski for the composite step.
/mixing-engineering-craftThe MIR metrics layer: onset, chroma, and the measurement vocabulary behind the alignment and key steps.
/mir-toolkit-referencePlugin-chain mixing. For finishing a lane or a composite once alignment holds.
/waves-mercury-mixerGrades wav quality against a bar. A programmatic gate to sit next to the audible review.
/synth-wav-graderDrum-pattern analysis reference. Reads what a kit lane is doing once split.
/midi-drum-analysisPairs drum and bass content. Directly relevant to the kick+bass summation test.
/drum-bass-matcherGrades a drum lane or loop for real signal, the guard against a full-size empty file.
/drum-loop-graderGrades a bass lane the same way. Pairs with the pYIN note-continuity check.
/bass-loop-graderDrum writing and realization craft, for anything downstream of the split.
/drum-craft · /drum-realizeThe worklist and delivery-proof engine from your handoff. A Drive link in the buyer thread is the only proof, and this skill enforces it. Runs daily.
/etsy-delivery-auditBuilds an audio reading end to end. The producer of the files that then need the loudness fix.
/chiron-audio-readingTranscribes local mp4 or wav with faster-whisper. Your transcript step for the readings.
/file-getterVoice generation, for any reading delivered as synthesized or cloned speech.
/speech-synth · /voice-cloneThe composite-alignment and tie-break logic (polarity, offset scoring, the keep-or-drop mix test) is not its own skill yet. jam-stem-rebuild is its closest existing home and already carries the Demucs-then-DrumSep decomposition.
The brief you pasted hands over a ready SKILL.md template for a drum-decomposition-and-composite-alignment skill. Turning that into a real skill with /skill-creator would make this whole pipeline one command next time instead of a hand-run sequence.
You asked for books and skills, so I did not touch the pipeline. These stay queued and ready when you say go.