Research map / Large language models / Retrieval and knowledge
Responding to the Call: Exploring Automatic Music Composition Using a Knowledge-Enhanced Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Call-Response Dataset (19,155 call-response pairs from POP909) and the Call-Response Generator (CRG), a Transformer seq2seq model that retrieves knowledge candidates from Pop1K7 by rhythm/melody/harmony similarity to generate musical responses to calls.
Problem
Automatic music composition models rarely produce call-and-response structure, so machine-composed music lacks the back-and-forth creative phrasing common in human music.
Approach
A compositional encoder-decoder conditioned on music knowledge: candidate phrases are selected from a large corpus by rhythm, melody and harmony similarity to the call, their hidden states are fused into the encoder output, and the model is trained with both ground-truth response loss and a loss to reconstruct candidate phrases.
Key results
- CRG achieves RQ 0.59, MQ 0.57, HQ 0.72 versus 0.30/0.26/0.52 for Music Transformer.
- In subjective survey (89 listeners), CRG creativity rating 3.91 vs 2.28 for Music Transformer, with statistically significant gains.
- Similar knowledge candidates give the largest boost over random or distinct knowledge (RQ 0.57 vs 0.55 random).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CRG's knowledge-enhanced training and response generation | CRG vs. MT · CRD test | RQ: 0.30 -> 0.59 | Helped |
| Use CRG's knowledge-enhanced training and response generation | CRG vs. MT · CRD test | MQ: 0.26 -> 0.57 | Helped |
| Add a knowledge mechanism and train with L1 rather than the L0-only baseline (ablation) | CRG-Base · CRD test | MQ: 0.29 -> 0.57 | Helped |
| Add L0 training before L1 (ablation) | CRG-K · CRD test | RQ: 0.48 -> 0.57 | Helped |
| Add L0 training before L1 (ablation) | CRG-K · CRD test | MQ: 0.57 -> 0.55 | Hurt |
| Use L2-only training instead of the L0-only baseline (ablation) | CRG-Base · CRD test | RQ: 0.44 -> 0.38 | Hurt |
| Use L2-only training instead of the L0-only baseline (ablation) | CRG-Base · CRD test | HQ: 0.57 -> 0.50 | Hurt |
| Combine L1 and L2 training with the knowledge mechanism (ablation) | CRG-KM1 · CRD test | RQ: 0.57 -> 0.59 | Helped |
| Combine L1 and L2 training with the knowledge mechanism (ablation) | CRG-KM1 · CRD test | MQ: 0.55 -> 0.57 | Helped |
| Combine L1 and L2 training with the knowledge mechanism (ablation) | CRG-KM1 · CRD test | HQ: 0.71 -> 0.72 | Helped |
Surprising findings
- Human-composed music was rated higher than CRG on quality, groove and coherence, yet CRG was rated more creative than human music.
Limitations
- Subjective evaluation relies on 89 social-media survey responses.
- Trade-off between music diversity and quality as the number of knowledge candidates grows.
- Response effects are not controllable in the current model.
Experimental setup
Datasets: POP909 Pop1K7
Benchmarks: Call-Response Dataset
Metrics: RQ MQ HQ n-gram diversity overall quality groove coherence creativity
Baselines: Music Transformer CP-Transformer HAT
Builds on: Compound Word Transformer Music Transformer POP909
Closest papers by meaning
- CMI-RewardBench: Evaluating Music Reward Models with Compositional Multimodal Instruction (ICML 2026 · similarity 0.88)
- MusicRL: Aligning Music Generation to Human Preferences (ICML 2024 · similarity 0.87)
- Compose with Me: Collaborative Music Inpainter for Symbolic Music Infilling (AAAI 2025 · similarity 0.87)
- LLark: A Multimodal Instruction-Following Language Model for Music (ICML 2024 · similarity 0.86)
- JEN-1 Composer: A Unified Framework for High-Fidelity Multi-Track Music Generation (AAAI 2025 · similarity 0.86)
- Adaptive Accompaniment with ReaLchords (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.