Research map / Vision-language and multimodal / Cross-modal learning
Unlocking Strong Supervision: A Data-Centric Study of General-Purpose Audio Pre-Training Methods
Xuanru Zhou, Yiwen Shao, Wei-Cheng Tseng, Dong Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UTS tag supervision from 400k clips instead of AudioSet tags from 2M clips | MTC pre-training · VoxCeleb2 linear probe (mean-pooling) | accuracy: 18.84 -> 37.10 | Improved |
| use UTS tag supervision from 400k clips instead of AudioSet tags from 2M clips | MTC pre-training · FSD-50k linear probe (mean-pooling) | mAP: 0.656 -> 0.459 | Hurt |
| use UTS-based MTC instead of AudioSet MTC for music question answering | MTC pre-training · AIR-Bench Music QA | GPT-4o score: 5.61 -> 6.16 | Improved |
| replace original captions with high-fidelity Qwen3-Omni-Captioner summaries | Captioning-scratch · AudioCaps and retrieval | Recall (retrieval): 49.2 -> 50.2 | Improved |
| replace original captions with high-fidelity captions | Contrastive-scratch · linear probing, attention pooling | mixed across tasks: FSD-50k 0.534 -> 0.514; VoxCeleb2 70.18 -> 71.63 | Mixed |
| add the tag multi-tag classification loss to the captioning objective (multi-task) (ablation) | Captioning-scratch · FSD-50k linear probe (mean-pooling) | mAP: 0.439 -> 0.485 | Improved |
| use parallel decoding objective instead of MTC (ablation) | UTS tag pre-training · AIR-Bench speech QA | accuracy: MTC gender 92.64 vs PAR 37.40 | Hurt |
| switch from mean-pooling to multi-head attention pooling (ablation) | Contrastive-scratch (Ours) · VoxCeleb2 linear probe | accuracy: 33.88 -> 71.63 | Improved |
| increase UTS vocabulary size beyond 1.5k to 2k (ablation) | MTC pre-training · linear probing and retrieval | normalized average performance | Hurt |
| increase UTS vocabulary size (ablation) | captioning · captioning | normalized average performance | No clear effect |
Closest papers by meaning
- AudioCALM: Continuous Autoregressive Language Modeling for Universal Audio Generation (NeurIPS 2026 · similarity 0.87)
- MATS: An Audio Language Model under Text-only Supervision (ICML 2025 · similarity 0.87)
- Massive Sound Embedding Benchmark (MSEB) (NeurIPS 2025 · similarity 0.87)
- CompA: Addressing the Gap in Compositional Reasoning in Audio-Language Models (ICLR 2024 · similarity 0.86)
- Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models (NeurIPS 2025 · similarity 0.86)
- A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization (NeurIPS 2026 · 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.
