PaperlayerView exact revision on arXiv

Source-linked AI summary

CloudEar: Integrating Perceptual, Diagnostic, and Cross-Modal Evidence for Music Evaluation

Qiqi He, Anqi Huang

arXiv:2610.11672v1eess.AS

TL;DR

Generated-music evaluation must address both perceptual and technical quality, and broad audio–text correspondence may miss fine-grained prompt attributes. CloudEar combines perceptual, diagnostic, and cross-modal experts in a pairwise framework; experiments report stronger performance than compared baselines across musicality and alignment evaluation.

  • Problem

    Generated-music evaluation must account for signal-level defects and perceptual quality, while broad audio–text similarity can miss fine-grained musical attributes requested by prompts.

    Source: §1 · p. 1

  • Method

    CloudEar is a pairwise framework combining perceptual quality modeling, signal-level and learned diagnostics, and genre- and instrument-based prompt-song alignment.

    Source: §1 · p. 1, §2.2 · p. 2, §2.3 · p. 2

  • Results

    CloudEar outperforms all compared baselines on the reported benchmarks, with higher musicality and alignment P-ACC and higher SRCC and LRCC on PAM-Music.

    Source: §3.2 · p. 2

  • Takeaways & Limitations

    Experiments on three human-annotated benchmarks show consistent improvements over existing methods for generated-song musicality and prompt alignment.

    Source: §4 · p. 2

Abstract

from arXiv · show

Evaluating generated music requires modeling perceptual quality, technical audio quality, and prompt alignment. Existing methods often overlook signal-level defects, produce compressed quality scores, and miss fine-grained musical attributes in text prompts. We propose CloudEar, a pairwise framework for evaluating musicality and prompt alignment. It comprises three components: i) a Perceptual Expert for song-quality assessment; ii) a Diagnostic Expert for learned and signal-level quality analysis; and iii) a Cross-modal Expert for prompt-song alignment based on musical attributes. Experiments show that CloudEar outperforms the compared baselines in pairwise preference accuracy and score correlation.

1. INTRODUCTION

Generated-music evaluation must account for both technical audio quality and musical aesthetics, while prompt alignment requires checking fine-grained musical attributes. CloudEar addresses these gaps with a pairwise evaluation framework, and experiments report improved musicality and alignment evaluation over existing methods.

1. INTRODUCTION

  • Existing evaluators may overlook signal-level defects and subtle perceptual differences, while regression models can compress score differences through central-tendency bias.Source: §1 · p. 1
  • General-purpose audio–text similarity can miss fine-grained requested attributes, including instrumentation, so high similarity does not ensure prompt adherence.Source: §1 · p. 1
  • CloudEar is a pairwise framework for evaluating musicality and prompt alignment.Source: §1 · p. 1
  • CloudEar consistently improves musicality and alignment evaluation over existing methods on rating and pairwise-preference benchmarks.Its three experts target song quality, signal-level quality, and prompt-song alignment.Source: §1 · p. 1

2. PROPOSED FRAMEWORK

CloudEar takes a prompt and two generated songs and predicts pairwise preferences for musicality and prompt alignment. Its Perceptual, Diagnostic, and Cross-modal Experts combine complementary quality and attribute evidence.

2. PROPOSED FRAMEWORK

  • Given a prompt and two generated songs, CloudEar predicts pairwise preferences in musicality and prompt alignment.Source: §2 · p. 1

2.2 Diagnostic Expert

  • The Diagnostic Expert fuses true-peak, clipping, crest-factor, and loudness descriptors with a HEAR quality score and the perceptual score to produce musicality scores.The fusion layer is trained on pairwise preference labels using binary cross-entropy, then selects the song with the higher score at inference.Source: §2.2 · p. 2

2.3 Cross-modal Expert

  • The Cross-modal Expert compares audio-derived genre and instrument labels with prompt attributes extracted by Qwen3-30B, then uses a Transformer to score alignment.It trains on pairwise preference probabilities for songs sharing a prompt and selects the higher-scoring song.Source: §2.3 · p. 2

3. EXPERIMENTS

Across the reported benchmarks, CloudEar outperforms the compared baselines, with stronger preference accuracy and rating correlations. The experiments evaluate both learned components and benchmark performance using human-annotated music data.

3.1 Datasets & Metrics

  • Training uses AIME, MusicArena, and about 7,000 hours of in-house human-rated music, with pairwise data from MusicEval and CMI-Pref; evaluation uses PAM-Music, CMI-Test, and MusicArena test data.The 1,340 MusicArena song pairs were generated in Jul.–Dec. 2025 with no training-data overlap; scalar ratings use SRCC and LRCC, and A/B preferences use P-ACC.Source: §3.1 · p. 2

3.2 Experimental Results

  • The full CloudEar outperforms reduced configurations on musicality; HEAR and DSP each improve over PE-only, and the alignment branch gains over PE-only.PE alone outperforms SongBench on CMI-Test musicality and remains competitive on MusicArena-Test.Source: §3.2 · p. 2
  • CloudEar outperforms all compared baselines on the reported benchmarks, with higher musicality and alignment P-ACC and higher SRCC and LRCC on PAM-Music.The table caption identifies comparison and ablation results for musicality and alignment; it does not provide cell values.Source: §3.2 · p. 2, Table 1 · p. 2

4. CONCLUSION

CloudEar integrates perceptual, diagnostic, and cross-modal experts to evaluate generated-song musicality and prompt alignment. Experiments on three human-annotated benchmarks show consistent improvements over existing methods.

4. CONCLUSION

  • CloudEar integrates perceptual, diagnostic, and cross-modal experts for generated-song musicality and prompt-alignment evaluation, with consistent improvements on three human-annotated benchmarks.Source: §4 · p. 2

5. AI USAGE STATEMENT

The authors used Qwen3-30B to extract structured genre and instrument attributes from input prompts and state that they take full responsibility for the paper's content.

  • Qwen3-30B extracts structured genre and instrument attributes from input prompts; the authors take full responsibility for the paper's content.Source: §5 · p. 3
Loading 2610.11672v1…