Research map / Label-efficient and robust learning / Knowledge distillation
MetricGAN-OKD: Multi-Metric Optimization of MetricGAN via Online Knowledge Distillation for Speech Enhancement
Wooseok Shin, Byung Hoon Lee, Jin Sob Kim, Hyun Joon Park, Sung Won Han
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 |
|---|---|---|---|
| one-to-one generator-metric OKD (MetricGAN-OKD) vs multi-metric baselines | MetricGAN multi-metric optimization · VoiceBank-DEMAND, PESQ+CSIG targets | PESQ: 3.20/3.24 vs 3.13-3.14 for baselines | Improved |
| one-to-one generator-metric OKD vs existing multi-metric methods | MetricGAN multi-metric optimization · Harvard Sentences (LE), SIIB/HASPI/ESTOI | SIIB: 79.40 vs 71.07 best baseline | Improved |
| spectrogram-level OKD with identical target metrics (ablation) | MetricGAN-OKD v2 (PESQ or CSIG only) · VoiceBank-DEMAND | CSIG: 4.26-4.27 vs 4.25 Multi-node and 4.23 single-metric | Improved |
| OKD weight alpha increased from 1 to 5 or 10 (ablation) | MetricGAN-OKD v1 · Harvard Sentences (LE) | score range across targets: Range SIIB 1.5 at alpha=1 vs 0.6 at alpha=5 vs 0.5 at alpha=10 | Improved |
| multi-metric OKD on four targets vs existing multi-metric methods | MetricGAN-OKD v2 · VoiceBank-DEMAND, four targets | COVL: 3.71-3.73 range vs 3.65-3.68 baselines for COVL; CBAK lower (3.07-3.14 in Table 2, 3.22-3.30 Table 3) | Mixed |
| OKD v2 vs multi-node MGAN on unseen mixed test set | MetricGAN-OKD v2 · Unseen LibriSpeech mix | COVL: 3.42 vs 3.26 | Improved |
| causal (unidirectional LSTM) MetricGAN-OKD vs non-MetricGAN causal models | MetricGAN-OKD v2 causal · VoiceBank-DEMAND | PESQ: 3.12 vs DEMUCS 2.93, CleanUNet 2.88 with 0.82M params vs 18.87M/39.77M | Mixed |
Closest papers by meaning
- FINALLY: fast and universal speech enhancement with studio-like quality (NeurIPS 2024 · similarity 0.82)
- DMOSpeech: Direct Metric Optimization via Distilled Diffusion Model in Zero-Shot Speech Synthesis (ICML 2025 · similarity 0.82)
- Self-Supervised Speech Quality Estimation and Enhancement Using Only Clean Speech (ICLR 2024 · similarity 0.81)
- Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement (NeurIPS 2026 · similarity 0.81)
- GenSE: Generative Speech Enhancement via Language Models using Hierarchical Modeling (ICLR 2025 · similarity 0.81)
- Complex-Cycle-Consistent Diffusion Model for Monaural Speech Enhancement (AAAI 2025 · similarity 0.80)
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.
