Research map / Learning theory and algorithms / Prediction and decision losses
Sample-Constrained Black Box Optimization for Audio Personalization
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.
Proposes OracleBO, a hybrid black-box optimization method that combines filter queries (user scores for full filters) and dimension queries (user's optimal per-frequency values) via batch q-EI and dimension-matched sampling on top of sparse GPR/ALEBO, for audio personalization.
Problem
Personalizing audio filters to maximize a user's unknown satisfaction function under a limited query budget, where users can be asked either to score full filters or to give optimal values for individual dimensions (frequencies).
Approach
Extends sparse (ALEBO-style) Bayesian optimization with a batch q-EI acquisition that proposes q candidate filters, then selects among them the one best matching the user's per-dimension optimal values (dimension-matched sampler), thereby combining both query types under one budget.
Key results
- With real volunteers rating audio on 0-10, OracleBO achieves an average of 3.3 points higher satisfaction than baseline within B=30 queries.
- On synthetic functions, the hybrid approach with dimension queries reduces regret relative to ALEBO(L=0), and the optimal number of dimension queries is around L=15 for a budget of 90.
- Choosing dimensions with largest variance (L_Top) achieves lower regret than random dimensions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine dimension queries with filter queries | OracleBO · Audio personalization | satisfaction: +3.3 points | Helped |
| Combine dimension queries with filter queries instead of using filter queries alone | OracleBO versus ALEBO(L=0) · Synthetic staircase satisfaction functions | regret | Helped |
| Increase dimension queries up to L=15 at fixed B=90 | OracleBO · Synthetic staircase satisfaction functions | regret | Helped |
| Increase dimension queries beyond L=15 at fixed B=90 | OracleBO · Synthetic staircase satisfaction functions | regret | Hurt |
| Select queried dimensions with largest function variance instead of random dimensions | OracleBO · Synthetic satisfaction functions | regret | Helped |
| Use OracleBO with Q_d=5 and Q_f=25 instead of the audiogram baseline | Audio personalization with hearing-loss distortion · Hearing-loss profile | satisfaction | Mixed |
| Use only Q_f=5 or Q_f=15 queries with OracleBO | Audio personalization with hearing-loss distortion · Hearing-loss profile | satisfaction | Hurt |
| Use OracleBO with Q_f=15 instead of the interpolated baseline | Audio personalization with random-filter distortion · Random distortion filter | satisfaction | Helped |
| Reduce the number of dimension queries | OracleBO · Audio personalization with random distortion filter | satisfaction | Hurt |
| Increase embedding dimension from d=4 to d=10 (ablation) | OracleBO · P1 regret (N=500, q=5, σ=1) | regret: 83 -> 148 | Hurt |
| Increase embedding dimension from d=4 to d=20 (ablation) | OracleBO · P1 regret (N=500, q=5, σ=1) | regret: 83 -> 477 | Hurt |
| Use q=2 and σ=0.2 instead of q=5 and σ=1 (ablation) | OracleBO · P1 regret (N=500, d=4) | regret: 83 -> 445 | Hurt |
Surprising findings
- Interpolated audiogram baseline performs well for hearing-loss simulation, so OracleBO only matched or improved it with enough filter queries (Q_f=25 with Q_d=5).
- Adding more dimension queries beyond a sweet spot (about 15) increases regret because fewer filter queries remain.
Limitations
- No theoretical convergence analysis; results are empirical only.
- Real-user experiment is tiny (3 volunteers without hearing loss, simulated hearing loss).
Experimental setup
Datasets: NHANES hearing-loss profiles synthetic satisfaction functions P1-P3 Branin Hartmann6 Rosenbrock
Benchmarks: P1-P3 staircase satisfaction functions Branin Hartmann6 Rosenbrock real volunteer audio personalization (3 users)
Metrics: regret user satisfaction score (0-10)
Baselines: ALEBO audiogram interpolation baseline corrupted audio
Builds on: ALEBO Expected Improvement Gaussian process regression q-EI
Closest papers by meaning
- BNMusic: Blending Environmental Noises into Personalized Music (NeurIPS 2025 · similarity 0.83)
- MusicRL: Aligning Music Generation to Human Preferences (ICML 2024 · similarity 0.83)
- DITTO: Diffusion Inference-Time T-Optimization for Music Generation (ICML 2024 · similarity 0.82)
- Kernel Learning for Sample Constrained Black-Box Optimization (AAAI 2025 · similarity 0.82)
- ETTA: Elucidating the Design Space of Text-to-Audio Models (ICML 2025 · similarity 0.82)
- Customized Condition Controllable Generation for Video Soundtrack (CVPR 2025 · similarity 0.82)
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.