Chapter 38 Exercises

This set is different from every other set in the book in one respect: some of these exercises will outlive the tools they mention. Work them as method training, not product training — the specific services you use this month will rebrand, merge, or fold into your DAW within a few years, and the audit loop you're drilling will still be the job. Do Part A first to load the frameworks, then Part B before Part C: the Listening Lab builds the evaluation ear that the DAW Lab assumes. The centerpiece is C1 + B1 together — the full AI audit and the blind master A/B from the Project Checkpoint — and they want a real evening each, not a coffee break. Standing laws apply everywhere: matched loudness for every comparison, full mix for every judgment, verdicts in writing, in band vocabulary. One more law specific to this chapter: every separated stem and generated file you make for these exercises is for private study — Chapter 36 still owns the question of what ships. Exercises marked ★ have worked answers in the back of the book (Answers to Selected Exercises).

Part A: Conceptual Understanding

A1. The two axes. ★ State the two questions this chapter measures every AI tool against. Then place these four tools on both axes, with one sentence of justification each: a stem separator, an automated mastering service, a chord generator, a full-track generative model.

A2. The solved one. Why does this chapter call stem separation "the genuinely solved one" while calling mixing assistants and mastering services "second opinions"? Your answer should name what separation replaces versus what assistants touch — the tedium/judgment distinction, in your own words.

A3. Artifact physics. Using the sidebar's mask intuition, explain in plain words why (a) cymbals come out of separation watery, (b) a faint vocal ghost haunts the extracted instrumental, and (c) reverb tails are the structurally unsolvable case. Three short paragraphs, no equations.

A4. The honest comparison frame. ★ Reproduce the three-way mastering comparison from memory as best you can: self-master, AI service, human engineer — across cost, speed, intent, EP context, and revision dialogue. Then answer the question the table is really asking: for your next release, which row, and why? (There is no wrong row. There is an unjustified row.)

A5. Steelman drill. Write the strongest single paragraph you can FOR training generative models on copyrighted catalogs without a license — then the strongest single paragraph AGAINST. Rule: each paragraph must be one a smart advocate of that side would endorse. If either paragraph contains a strawman, rewrite it. Finish with two sentences on which argument moves you more and what evidence would change your mind.

A6. The consent lines. A singer asks you to produce a demo using (a) a clone of her own voice for harmony drafts, (b) a licensed voice model from a marketplace, (c) a clone of a famous artist "so the label gets the vibe." Give your answer to each, with the reasoning the chapter gave — and for (c), draft the two-sentence reply you'd actually send, professional but unambiguous.

A7. Why the ear survives. Summarize the chapter's two structural arguments for why trained evaluation keeps its value as tools improve (the outcomes-in-general vs outcomes-in-particular argument, and the generation-vs-discrimination bottleneck). Then connect each to a specific chapter of this book that built the relevant skill.

A8. Provenance hygiene. Your distributor's upload form asks: "Does this release contain AI-generated content?" For a track where you used stem separation to study a reference, an assistant EQ starting point you heavily edited, and a synthetic guide vocal that was replaced by a human singer — what's the honest answer, and what would your session's provenance log need to say to back it up? One paragraph.

Part B: 🎧 Listening Lab

Ear training never stops, and this chapter's listening skill — auditing machine output — is the one your career will lean on hardest. Log everything in your Chapter 4 journal: date, system, verdicts in band vocabulary.

B1. The blind master A/B (the big one). ★ Execute the Project Checkpoint experiment in full: your Chapter 32 master vs one automated mastering service's master of the same Chapter 31 print. Protocol, no shortcuts: measure integrated LUFS on both; trim the hotter file down to match; have someone else (or a randomizing playlist) label them X and Y; full-song listens on your monitors, earbuds, and a phone speaker; score each on low end, vocal presence, top end, punch/density, the quietest section, and "which is the record." Unblind only after all scores are written. Deliverable: the scorecard, plus three sentences on what the service matched, what it missed, and what (if anything) survives a matched-loudness audition into your master.

B2. Grade your prophecy. Before unblinding B1, retrieve your Productive Struggle prediction note. After unblinding, grade it category by category. Where you predicted wrong, write one sentence on why — wrong about the tool, or wrong about your own master? That distinction is the most personal ear-training data this book can generate.

B3. Spot-the-generated drill — honestly framed. Assemble five pairs of short clips: one human-made commercial track and one fully generated track per pair (generated examples are abundant as of this writing; a friend can source and shuffle them if you want it blind). Identify which is which, and — this is the actual exercise — write down the tell you used, in band and craft vocabulary: consonant clarity, cymbal smear, low-end focus, arrangement intent, lyric coherence. Honest framing, read it twice: the tells you find today will fade as models improve; the method — listening for intent and naming evidence — is the durable skill. Re-run this drill in a year and expect a worse score with better ears.

B4. Stem-separation artifact hunt. ★ Separate one commercial track you know deeply (private study only). Solo each stem at healthy level and log every artifact you can name: watery cymbals, vocal ghost in the instrumental, dulled transients, reverb clinging to the wrong stem, "other" stem mush. Then the diagnostic step: for each artifact, note WHERE it lives in band vocabulary (the ghost sits 1–4 kHz; the swirl lives above 6 kHz). Finish by muting and unmuting stems against the full mix — at what blend level does each artifact disappear under cover? That threshold is your practical usability map.

B5. Audit the assistant. Run an assistant EQ or assistant chain on one track from your Chapter 22-era mix session. Do not accept or reject yet. Instead, read the suggestion and translate every node into Chapter 30 vocabulary aloud: "it cut 3 dB at 250 — it thinks the track is muddy." THEN A/B each node at matched loudness in the full mix and rule on it: confirmed, rejected, or modified. Deliverable: the suggestion's hit rate, and one sentence on whether its misses shared a pattern (they usually do: genre misread, or solo-context logic in a full-mix world).

B6. The regression-to-the-mean listen. Pick one beloved record that violates its genre's conventions hard (you profiled candidates in Chapter 18) and one genre-typical track from the same era and style. List three deviations the beloved record makes that an averaging system would "correct" — an arrangement drop where the genre never drops, a vocal drier or darker than convention, a low end that breaks the tonal-balance window. For each: what would the correction sound like, and what would die with it? Then the uncomfortable final question, answered honestly in your journal: which of YOUR track's deviations would survive the same treatment, and do you love them enough to defend them in a blind test? This is the identity-cost argument made audible — and personal.

B7. The louder lie, replicated. Take B1's two masters and deliberately A/B them UNMATCHED — the service's master at its delivered level against yours. Note your gut preference. Then re-match loudness and A/B again. Write down what changed in your preference and which bands drove the flip. You proved this bias in Chapter 22 with EQ moves; feeling it flip your verdict on an entire master is the version you'll never forget.

B8. The restoration listen. Listen to the 2023 Beatles release "Now and Then" alongside one mid-1990s Beatles reconstruction ("Free as a Bird" or "Real Love"). Focus on the lead vocal: presence, intimacy, freedom from artifacts of its source. Write one paragraph on what the demixing technology bought that 1995's tools couldn't — and one paragraph on everything in the record that remained stubbornly human work. Case study 1 is the companion read.

Part C: 🎛️ DAW Lab

C1. The full AI audit worksheet. ★ The Project Checkpoint's centerpiece, executed in writing. Build a table: one row per stage of your track's production (capture plan, recording, editing, beat programming, sound design, arrangement, polish, each mix pass, master). Columns: Could a current tool help here? / Can I evaluate its output? / Tedium or judgment? / Identity cost? / Verdict (adopt, second-opinion, not yet, never). Fill every row with a real answer, not a vibe. Deliverables: the table, plus a three-line summary — your top adoption, your hardest "never," and the "not yet" whose ear-training you'll prioritize.

C2. The separation experiment. Take your OWN Chapter 31 mix print (or a collaborator's, with permission) and run it through a stem separator. In your DAW: line the stems up against the original multitracks and compare — where did separation lose information your session never had to guess about? Then the constructive half: build a 16-bar remix bed from the separated stems alone (new drums under the separated vocal, for instance). Log which artifacts forced arrangement decisions — the practical lesson of working FROM stems instead of sessions.

C3. The assistant-vs-your-EQ comparison. Duplicate one track you mixed by hand in Chapter 22 (vocal or bass is ideal). On the duplicate, run an assistant's suggested chain unedited. Level-match the two versions, blind-shuffle them in the full mix, and pick a winner per listening pass: clarity, tone, fit with neighbors. Then build the hybrid: start from YOUR version and audition only the assistant nodes you ruled "confirmed" in B5. Deliverable: which of the three versions ships, and the one-sentence reason — this is the starting-point-vs-ending-point doctrine, executed.

C4. Sketch-to-human finish. Generate raw material once: an 8-bar progression from a chord tool, or a short generative sketch in your genre. Now produce a 30–60 second piece where NONE of the generated audio survives — you may keep ideas (a chord color, a rhythmic gesture, an arrangement shape), but every sounding element is re-played, re-programmed, or re-designed by you (Chapter 9 humanization, Chapter 14 patch design). No tool access? Run the identical workflow from a reference track's chord chart — which is the point: the workflow is "steal the idea, perform the record," and it predates the robots.

C5. The provenance log template. Open your Chapter 19 session-notes document and add an AI-use section: date, tool category, what it touched, what survived into the record, cleared-or-study-only status. Backfill it honestly for every exercise in this set, then for your whole track. Ten minutes now; the habit is the deliverable — from this session forward, the log is part of hygiene.

C6. The topline draft workflow (optional gear). If you have access to a voice model (your own voice, or a licensed one): draft a topline over an instrumental — melody, phrasing, breath placement — then write the handoff document you'd send a real singer: range, the phrases that matter, where the draft is wrong on purpose. No access? Hum the draft into your phone and extract MIDI. Either way, the deliverable is the handoff doc: proof the draft is scratch paper serving a human performance, not a substitute for one.

C7. The EP-context experiment. Automated mastering's structural blind spot is the set — it masters one file at a time, blind to siblings. Test it: take two of your own tracks (or your track plus a stylistically adjacent older one), run each through the same service independently, and line the returns up back-to-back in your DAW against your own Chapter 32-style sequence of the same pair. Measure the integrated LUFS of each file and note the deltas; then listen across the junction both ways at matched playback level. Questions to answer in writing: do the automated masters agree with each other on tonal balance, or did each track get pulled toward a different genre guess? Does the louder-to-quieter handoff feel intentional or accidental? What did your human sequencing decide (Chapter 32's level map and spacing) that no track-at-a-time process could? One paragraph of findings — this is the table's "hears your EP as a set?" row, made audible.

Part D: Synthesis

D1. The ethics debate (structured). The scenario: a producer releases a tribute single using a voice model of a beloved singer who died ten years ago, with the estate's permission but with no disclosure to listeners. Stage the debate — with a collaborator, or as a written dialogue: Round 1, argue it's ethical (consent exists, tribute intent, estates license likenesses all the time). Round 2, argue it isn't (the audience's trust is undisclosed-synthetic; the dead can't consent to new "performances"; precedent risk). Round 3: write the disclosure policy YOU would require to sign off on the release — and state whether any policy suffices. There's no answer key. There's a position you can defend, or there isn't.

D2. The five-year memo, yours. The chapter's speculation section modeled the form; now write your own one-page memo: how AI probably changes production in YOUR genre in five years — three "probably" claims, two "maybe" claims, one "what stays" claim, every verb honest. Date it. Calendar a reminder to re-read it in one year and grade it in the margins. (This is forecasting practice, and forecasting practice is humility practice.)

D3. The training-data position paper. One page: your position on training generative music models on copyrighted catalogs. Required structure: steelman the opposing side FIRST (a full paragraph), then your position, then the policy you'd actually vote for (licensing pools? opt-in registries? status quo?), then the one piece of evidence that would change your mind. Cite the chapter's arguments where you lean on them.

D4. Your studio's AI policy. Draft the one-page policy for your own work, suitable for showing a collaborator or client: what you use freely (and why it's tedium), what you use as second opinion (and the audit ritual attached), what you disclose and where, and what you won't touch (and whether that's ethics, identity, or both). Date it, revisit it annually alongside your D2 memo, and expect the "freely" list to grow while the "never" list holds — that asymmetry is what a principled policy looks like. Append it to your Chapter 5 production manifesto — it's the same document, grown up.

Part M: Mixed Practice

Interleaving is the point — these reach back across the whole book on purpose.

M1. (Ch 22 / Ch 4 / Ch 30) An automated master adds a high shelf and your snare suddenly reads "harsh" on earbuds but fine on monitors. Name the likely band, the Chapter 30 symptom category, the 60-second check, and two fixes at different stages (master-stage and mix-stage). ★

M2. (Ch 33) Your master measures -9 integrated LUFS; the AI master of the same print measures -8. A streaming platform normalizes to roughly -14 as of this writing. What happens to each file at playback, what survives the turn-down, and which dynamics decision actually matters now that the war is over?

M3. (Ch 36 / Ch 13) A client hands you a "clean a-cappella" they extracted from a 1990s R&B hit and asks you to flip it for a release. List the rights in play, who controls each (in role terms), and the two legitimate paths forward — plus the one-sentence explanation you'd give the client for why the separator changed nothing legally.

M4. (Ch 9 / Ch 13) You extract MIDI from a drum loop whose feel you love, drop it on your own kit, and it sounds dead. Diagnose: name three things the extraction likely lost or flattened, and the Chapter 9/13 moves that restore intention (not randomness) to the pattern.

M5. (Ch 30 / Ch 32 / Ch 19) A bandmate returns from an automated service with a master that "sounds amazing" and wants to ship tonight. Design the 30-minute verification session you'd run together: the measurements, the matched-loudness listens, the three systems, the scorecard categories — and the feedback-literacy framing (Chapter 19) you'd use to deliver a "not yet" without torching the relationship. Bonus point: name the one outcome of the session in which the right answer IS to ship the automated master tonight, and what that outcome would tell you about the band's mix.