Chapter 38 Further Reading

Everything below is real as of this writing — and this chapter's reading list ages faster than any other in the book, so two standing rules: search by title and author rather than trusting any link, and prefer sources that teach the durable layer (how separation works, how the legal questions are structured, how to evaluate output) over tool reviews, which rot in months. A warning specific to this topic: AI-music coverage online is dominated by two equally useless genres — doom content and demo-reel hype, both optimized for clicks rather than craft. The sources below were chosen because they commit neither sin.

Beginner

  • Now and Then — The Last Beatles Song (official making-of short film, 2023). Twelve minutes, free to watch online, directed by Oliver Murray. The single best on-ramp to this chapter's tool-versus-replacement question: you see the cassette, hear the isolated vocal, and watch every human decision that followed. Pair it with this chapter's case study 1 and the song itself.
  • Your own DAW's assistive features, documentation first. Most major DAWs and plugin suites now ship assistant and separation features with genuinely good explanatory documentation. Reading what the manufacturer says the tool does — before using it — is beginner-level provenance hygiene and surprisingly rare practice.
  • Mainstream explainers on machine learning for audio. Search for introductory articles and videos on "source separation" and "how AI mastering works" from established audio-education outlets; the good ones map directly onto this chapter's sidebar. Skip anything promising "the death of" or "the end of" anything.

Intermediate

  • Sound on Sound magazine's coverage of AI production tools. The long-running industry magazine reviews assistant, separation, and mastering tools with exactly this chapter's temperament: tested claims, named limitations, no hype. Their workshop pieces on automated mastering services — comparing them against human engineers on real material — are the published version of your B1 exercise.
  • The Recording Academy's published AI guidelines. Short, primary-source, and the cleanest statement of the human-authorship line as awards bodies draw it (as of this writing). Reading the actual text beats every paraphrase of it.
  • Reputable music-industry trade coverage of the training-data litigation — Billboard's legal desk and the industry newsletter Music Ally both track the suits, settlements, and licensing deals in categories rather than headlines. Follow the litigation through them rather than through general news, which reliably garbles the two-copyrights distinctions you learned in Chapter 36.

Advanced

  • The public source-separation research literature. The field's landmark open papers are genuinely readable with this book's vocabulary plus patience: the Spleeter paper (Hennequin and colleagues, Journal of Open Source Software, 2020) describes the spectrogram-mask approach from the sidebar; the Demucs line of papers (Défossez and colleagues) shows the move to waveform-domain models and why artifacts shrank; Open-Unmix (Stöter and colleagues, JOSS, 2019) is the friendliest reference implementation. All are free, citable, and honest about failure modes in ways product pages never are.
  • The MUSDB18 benchmark and the source-separation evaluation campaigns. The standard public dataset and the community evaluations built on it document, in numbers, exactly how separation quality improved year over year — and what "state of the art" actually measures. Useful inoculation against marketing claims.
  • ISMIR proceedings (the International Society for Music Information Retrieval). The field's main conference publishes open-access papers on everything this chapter touched — separation, transcription, generation, evaluation. Browsing one year's proceedings calibrates you on what's solved, what's close, and what's still research fantasy.
  • The US Copyright Office's reports on artificial intelligence and copyright. Primary source for the human-authorship threshold and the office's evolving analysis of training and infringement questions (as of this writing, issued in parts). Dense but quotable, and the document your entertainment lawyer will be reasoning from.

For Educators

  • Structure the ethics material as debates, not lectures. Exercise D1 (the disclosed-estate-clone scenario) and D3 (the training-data position paper) are designed for it: require the steelman first, grade the quality of the opposing argument as heavily as the position. The goal is students who can hold a defensible position in a room of professionals, not students who memorized this year's lawsuit names.
  • The Copyright Office reports and the Recording Academy guidelines make excellent primary-source close-reading assignments — short enough for a seminar, current enough to feel alive, and they teach the hedging habit: have students date every claim they extract.
  • Run B1 (the blind master A/B) as a class lab. One mix print, one automated service, one student-built blind protocol, whole-class scorecard. It teaches matched-loudness discipline, evaluation vocabulary, and intellectual honesty in a single session — and the years-over-years drift in the scorecard, if you keep records, becomes its own longitudinal lesson about the landscape.