Research map / Label-efficient and robust learning / Knowledge distillation
Listen like a Teacher: Mitigating Whisper Hallucinations Using Adaptive Layer Attention and Knowledge Distillation
Kumud Tripathi, Aditya Srinivas Menon, Aman Gaurav, Raj Prakash Gohil, Pankaj Wasnik
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 |
|---|---|---|---|
| fine-tune Whisper-small on noisy data (Baseline-2 vs pretrained) | Whisper-small · Hindi Kathbath avg | WER (%): 141.08 -> 21.44 | Improved |
| add Adaptive Layer Attention to fine-tuned Whisper-small | Baseline-2 · Hindi Kathbath avg | WER (%): 21.44 -> 19.64 | Improved |
| add Adaptive Layer Attention | Baseline-2 · English LibriSpeech avg | WER (%): 12.46 -> 9.68 | Improved |
| add multi-objective KD on top of ALA | W-ALA · English LibriSpeech avg | WER (%): 9.68 -> 8.56 | Improved |
| use MOKD instead of Distil-Whisper distillation | Whisper-small · Hindi Kathbath avg | WER (%): 51.39 -> 18.61 | Improved |
| block-wise mean representations + MHA instead of weighted sum over layers (ablation) | ALA · Hindi Kathbath SNR -10 | WER (%): 75.85 -> 40.74 | Improved |
| make encoder layers trainable during MHA (ablation) | ALA with MHA over all layers · Hindi Kathbath SNR -10 | WER (%): 52.73 -> 45.12 | Improved |
| group layers into blocks with mean pooling instead of MHA on all layers (ablation) | ALA · Hindi Kathbath SNR -10 | WER (%): 45.12 -> 40.74 | Improved |
| add cosine loss on final layers only instead of all layers (ablation) | KD · Hindi Kathbath SNR -10 | WER (%): 45.61 -> 42.97 | Improved |
| add IRLE loss on final representations (ablation) | KD · Hindi Kathbath SNR -10 | WER (%): 42.77 -> 61.14 | Hurt |
| add OT loss on final representations (ablation) | KD · Hindi Kathbath SNR -10 | WER (%): 42.77 -> 46.53 | Hurt |
| add KL on logits (ablation) | KD · Hindi Kathbath SNR -10 | WER (%): 42.77 -> 46.48 | Hurt |
Closest papers by meaning
- MoLE:Decoding by Mixture of Layer Experts Alleviates Hallucination in Large Vision-Language Models (AAAI 2025 · similarity 0.88)
- Sample Transform Cost-Based Training-Free Hallucination Detector for Large Language Models (NeurIPS 2026 · similarity 0.88)
- The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs? (NeurIPS 2025 · similarity 0.88)
- Visual Hallucination Elevates Speech Recognition (AAAI 2024 · similarity 0.88)
- MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation (ICLR 2025 · similarity 0.88)
- SVHalluc: Benchmarking Speech-Vision Hallucination in Audio-Visual Large Language Models (CVPR 2026 · similarity 0.87)
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.
