PaperlayerView exact revision on arXiv

Source-linked AI summary

DJustify: Reason-and-Verify DJ Planning for Pop Transitions and Story-Driven Sets

Wei-Han Hsu, Li-Jie Lin, Po-Hsuan Lai, Po-Hsiang Huang, Jeng-Yue Liu, Li Su, Yi-Hsuan Yang

arXiv:2610.33434v1cs.SD

TL;DR

Vocal-rich pop makes automatic DJ mixing difficult because track choice, cue placement, and transition technique are coupled. DJustify uses an LLM to plan these choices together and checks rendered audio to guide repairs. Its transitions outperform prior automatic systems, while story-driven sets improve coherence over a no-story ablation and show no detectable rating gap from a human DJ.

  • Problem

    In vocal-rich pop, track choice, cue placement, and transition technique are coupled, while most automatic DJ systems target more regular EDM.

    Source: §1 · p. 1

  • Method

    DJustify samples LLM plans that jointly choose tracks, cues, and techniques, then uses deterministic acoustic checks on rendered audio to guide repair.

    Source: §1 · p. 1, §3.2 · p. 3

  • Results

    DJustify transitions outperform prior automatic systems, and Story improves coherence over No-story while showing no detectable story or transition rating gap from a human DJ.

    Source: §5 · p. 4, §4.4 · p. 4

  • Takeaways & Limitations

    Separating plan generation from rendered-audio verification is a practical design for AI-DJ systems that preserves multiple alternatives rather than prescribing one fixed mix.

    Source: §5 · p. 4

Abstract

from arXiv · show

Automatic DJ systems mostly target electronic dance music, whose regular structure makes rule-based mixing relatively easier. Vocal-rich popular music is harder: which track to play next, where to cue it, and which transition to use are tied together, so fixed rules handle them poorly. We present DJustify, a pop AI-DJ that puts an LLM in a reason--propose--verify--repair loop and lets it choose the next track, cue points, and mixing technique together. The LLM samples several transition plans; an external acoustic checker then scores the rendered audio and triggers repair when a check fails. Sampling keeps several candidate transitions per track pair instead of one fixed mix. In blind listening tests our transitions beat Mosaikbox, a reimplemented crossfade, and a cue baseline that reuses the crossfade's cues; that last comparison shows most of the margin comes from cue placement. The same planner also picks and orders a set along an emotion arc: this improves story coherence over a no-story ablation, and within a set story and transition ratings show no measurable gap from a human DJ. The checker mainly adds reliability, not higher average quality: it repairs rare but audible errors such as cut vocals, while the planner stays non-deterministic.

1. INTRODUCTION

Vocal-rich pop makes automatic DJ mixing difficult because track choice, cue placement, and technique are coupled. DJustify addresses this with sampled LLM plans that are checked and repaired on rendered audio.

1. INTRODUCTION

  • Pop transition quality depends jointly on the next track, cue points, and mixing technique, unlike the more regular structures that make fixed-rule EDM mixing easier.A technique that works for one cue pair may fail for another, and changing the next track changes which cues and techniques are feasible.Source: §1 · p. 1
  • Existing rule-based systems make coupled decisions deterministically and do not revise them after hearing the rendered mix.Sampling an LLM can produce several context-dependent plans rather than a single fixed answer.Source: §1 · p. 1
  • DJustify renders each LLM plan, uses deterministic acoustic gates to identify failures and suggest changes, and lets the planner repair it.The rules verify and bound valid transitions rather than deciding the output; the same loop extends to story-driven sets.Source: §1 · p. 1
  • The system significantly outperforms prior automatic systems, including a Spotify-style crossfade, while blind tests find no detectable story or transition rating gap from a human DJ for story-driven sets.A controlled cue baseline isolates the importance of cue-point selection.Source: §1 · p. 1

2. RELATED WORK

Prior automatic DJ work largely targets EDM and treats selection, cueing, and synthesis separately, while LLM music work has not addressed verification of rendered transitions. DJustify joins these decisions in a reason–verify loop and plans sets around an emotional arc.

2. RELATED WORK

  • Automatic DJ systems range from tempo- and energy-based sequencing to constrained cueing and learned transition generation, but most target EDM.Vocal handling has focused on stem separation or overlap detection rather than cue timing that preserves sung phrases; DJustify does not compare against Vande Veire’s full drum-and-bass pipeline.Source: §2 · p. 2
  • Prior systems optimize track selection, cueing, and synthesis in isolation, leaving play order without a planned emotional shape; DJustify combines them and plans an emotional arc.The paper notes that prior automatic DJ mixing systems do not plan play order around such an arc.Source: §2 · p. 2
  • LLM research includes reasoning, self-revision, and music agents, but self-generated critique can fail without external feedback and related music-agent work reasons over symbolic labels rather than rendered audio.Chang et al. [17] use chain-of-thought to fuse music-information-retrieval outputs for chord recognition over symbolic labels.Source: §2 · p. 2

3. METHOD

DJustify uses one LLM and a shared verified-planning loop for transitions and story-driven sets. It preprocesses tracks into features and cue anchors, then checks rendered plans and repairs failures.

3. METHOD

  • A single instruction-tuned LLM drives both transition planning and story-set construction, with the task prompt distinguishing the two levels.The system runs locally on one 48 GB NVIDIA RTX 6000 Ada GPU with a 4-bit-quantized model and no closed-source APIs.Source: §3 · p. 2

3.1. Song Preprocessing

  • Preprocessing estimates beat grids, structure, chords, singing-voice activity, and valence/arousal, then restricts cues to musically strong on-grid anchors.Anchors include phrase ends, section boundaries, harmonic restarts, and low-frequency onsets.Source: §3.1 · p. 2

3.2. Verified DJ Transition Planning

  • For each outgoing track, the planner jointly chooses an incoming track, cue-out and cue-in, overlap length, and transition technique from candidate options.Techniques range from long-overlap blends to abrupt effect-assisted cuts; rendered-audio verification is needed because symbolic anchors resolve to seconds only at render time.Source: §3.2 · p. 2
  • Deterministic acoustic gates inspect rendered audio and return actionable repairs; the planner acts for up to four rounds, after which a still-failing plan is flagged.The checker covers four gate families; the ablations remove reasoning, checker-driven repair, or both.Source: §3.2 · p. 3, Table 1 · p. 3

3.3. Story-Driven DJ Set

  • For story-driven sets, the LLM selects and orders songs into an emotional arc under tempo and key gates, then verifies each junction with the transition loop.The No-story ablation retains the gates and checker but selects each next track by acoustic compatibility alone; lyric-hook landing points also inform cue placement.Source: §3.3 · p. 3

4. EXPERIMENTS AND RESULTS

Across transition, diversity, reliability, and set-level tests, DJustify performs better than automatic baselines on key comparisons while retaining multiple transition options. Checker-driven repair corrects specific audible errors, but does not produce a significant overall rating gain over the unverified variants.

4.1. Setup

  • The study uses blind listening tests and an objective ablation, with fixed track pairs in Test1 and three-minute, prompt-based sets in Test2.The tests had 103 and 99 respondents, respectively; the objective ablation measures renderability and transition diversity.Source: §4.1 · p. 3

4.2. Transition Quality and Cue Placement

  • 5.6 mean transition rating for Ours exceeds Mosaikbox (4.0), Bittner’17/Spotify (4.7), and Ours (Bittner cues) (4.8), but remains below the human DJ (6.7).All four comparisons remain significant under Holm correction.Source: Figure 2 · p. 3, §4.2 · p. 3
  • Replacing the planner’s vocal-aware cues with Bittner’s detector cues lowers the score to 4.8, and repair does not close the gap.The cue baseline matches the planner’s cues on only 1 of 12 pairs.Source: §4.2 · pp. 3–4

4.3. Diversity and Reliability under Verification

  • The full system returns a median 6.5 distinct rendered transitions per pair, versus one for the deterministic reference.Reasoning increases median distinct proposals from three to five (p = .004); with reasoning enabled, repair’s increase from five to 6.5 is not significant (p = .054).Source: §4.3 · p. 4, Figure 3 · p. 4
  • Renderability rises from 91% without reasoning or checker repair to 99% with both, but listening ratings show no significant overall difference among variants.Checker-driven repair nevertheless removes mid-phrase outgoing-vocal cuts on two seeds; these rare rescues do not establish an overall perceptual-quality gain.Source: §4.3 · p. 4

4.4. Story-Driven Sets

  • Story improves story coherence over No-story by roughly 1.9 points (p = .002), while Story and the human DJ are indistinguishable on story coherence (p = .95).Story also has no significant transition-quality difference from the human DJ (p = .42); the Story–No-story contrast reflects selection and ordering together, including the opener policy.Source: §4.4 · p. 4, Figure 4 · p. 4

5. CONCLUSION

DJustify jointly plans track selection, cue placement, and transition technique in a reason–propose–verify–repair loop, while sampling multiple plans for each context. Blind listening tests find that its transitions outperform prior automatic systems, while story-driven sets show no detectable gap from a human DJ in story or transition ratings.

  • Blind listening tests find that DJustify transitions outperform prior automatic systems but remain below a human DJ, while story-driven sets show no detectable gap from a human DJ in story or transition ratings.Source: §5 · p. 4
  • Verification improves reliability rather than average quality, repairing rare audible errors such as cut vocals.Source: §5 · p. 4
  • Sampling preserves multiple rendered alternatives per track pair, yielding constrained diversity rather than a single deterministic mix.Source: §5 · p. 4
Loading 2610.33434v1…