Research map / Large language models / Language, safety and interpretability
Beyond Transcription: Mechanistic Interpretability in ASR
Neta Glazer, Yael Segal-Feldman, Hilit Segev, Aviv Shamsian, Asaf Buchnick, Gill Hetz, Ethan Fetaya, Joseph Keshet, Aviv Navon
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 |
|---|---|---|---|
| probe internal encoder representation instead of reading the text output | Qwen2-Audio · speaker gender, LibriSpeech | accuracy (%): 87.8 -> 94.6 | Improved |
| patch cross-attention at decoder layer 23 (ablation) | Whisper · repetition hallucinations | cases resolved (%): 76% | Improved |
| patch decoder self-attention or feed-forward layers (ablation) | Whisper · repetition hallucinations | repetition rate | No clear effect |
| patch cross-attention head 13 in layer 18 alone (ablation) | Whisper · repetition hallucinations | repetition suppressed (%): 78.1% | Improved |
| patch encoder components with white noise (ablation) | Whisper · acoustic-contextual errors | restored accuracy: 130/153 (85.0%) restored | Improved |
| patch decoder components with white noise (ablation) | Qwen2-Audio · acoustic-contextual errors | restored accuracy: 147/251 (58.6%) restored vs 171/251 via encoder | Improved |
| apply final encoder layer normalization before passing intermediate layers to decoder (Encoder Lens) (ablation) | Whisper · encoder lens | output coherence | Improved |
| probe decoder residual stream at later layers for hallucination prediction | Whisper · LibriSpeech test-clean | probe test accuracy: 0.622 (L5) -> 0.934 (L22) | Improved |
| probe decoder residual stream for hallucination on LibriSpeech | Qwen2-Audio vs Whisper · LibriSpeech | probe accuracy (%): 93.4 -> 70.2 | Hurt |
Closest papers by meaning
- Recording for Eyes, Not Echoing to Ears: Contextualized Spoken-to-Written Conversion of ASR Transcripts (AAAI 2025 · similarity 0.85)
- Towards Global-level Mechanistic Interpretability: A Perspective of Modular Circuits of Large Language Models (ICML 2025 · similarity 0.84)
- Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders (NeurIPS 2025 · similarity 0.84)
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.84)
- Self-Taught Recognizer: Toward Unsupervised Adaptation for Speech Foundation Models (NeurIPS 2024 · similarity 0.84)
- LLM Layers Immediately Correct Each Other (NeurIPS 2025 · 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.
