Research map / Image generation / Image restoration
Query-Based Asymmetric Modeling with Decoupled Input–Output Rates for Speech Restoration
Ui-Hyeop Shin, Jaehyun Ko, Woocheol Jeong, Hyung-Min Park
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 |
|---|---|---|---|
| add decoder with cross-attention (MHCA) to encoder-only model (ablation) | TF-Restormer · VCTK SSR 8->16 kHz | LSD: 2.12 -> 0.89 | Improved |
| add MHCA cross-attention in decoder instead of self-attention only (ablation) | TF-Restormer · VCTK SSR 8->44.1 kHz | LSD: 1.35 -> 1.01 | Improved |
| use size-matched small encoder-decoder (10.9M) instead of encoder-only (11.6M) (ablation) | TF-Restormer · VCTK SSR 8->16 kHz | LSD: 2.12 -> 1.36 | Improved |
| use encoder-decoder instead of encoder-only (ablation) | TF-Restormer · VCTK SSR 8->44.1 kHz | MACs (G): 415.1 -> 308.4 | Improved |
| use shared SFI-STFT discriminator instead of separate rate-specific discriminators (ablation) | TF-Restormer · VCTK+DEMAND DN | PESQ: 3.28 -> 3.41 | Improved |
| add shared frequency projection layer (ablation) | TF-Restormer · UNIVERSE GSR | UTMOS: 3.90 -> 4.10 | Improved |
| use separate (per-module) frequency projections instead of shared (ablation) | TF-Restormer · VCTK SSR 8->16 kHz | PESQ: 3.70 -> 3.55 | Hurt |
| add spectral loss on top of perceptual loss alone (ablation) | TF-Restormer pretraining · UNIVERSE GSR | PESQ: 1.85 -> 2.23 | Improved |
| replace l1 with l2 on complex STFT components (ablation) | TF-Restormer pretraining · UNIVERSE GSR | PESQ: 2.23 -> 2.21 | Hurt |
| use plain log1p (w=1) instead of l1 (ablation) | TF-Restormer pretraining · UNIVERSE GSR | PESQ: 2.23 -> 2.25 | No clear effect |
| use adaptive w_tf scaled log-spectral loss instead of fixed scale w=1e-3 (ablation) | TF-Restormer pretraining · VCTK SSR 8->16 kHz | MCD: 1.37 -> 1.29 | Improved |
| use overly small scale w=1e-4 in s-log1p (ablation) | TF-Restormer pretraining · VCTK SSR 8->16 kHz | MCD: 1.37 -> 2.43 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- FINALLY — FINALLY: fast and universal speech enhancement with studio-like quality (NeurIPS 2024) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- RestoreGrad: Signal Restoration Using Conditional Denoising Diffusion Models with Jointly Learned Prior (ICML 2025 · similarity 0.86)
- TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation (ICLR 2025 · similarity 0.86)
- Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (NeurIPS 2024 · similarity 0.86)
- Predictive–Generative Drift Decomposition for Speech Enhancement and Separation (NeurIPS 2026 · similarity 0.85)
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.85)
- Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement (NeurIPS 2026 · similarity 0.84)
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.
