Research map / Architectures and efficiency / Attention and transformers
Speech-Aware Long Context Pruning and Integration for Contextualized Automatic Speech Recognition
Yiming Rong, Yixin Zhang, Ziyi Wang, Deyang Jiang, Yunlong Zhao, Haoran Wu, Shiyu Zhou, Bo Xu
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 contextual keywords (5 slides) with SAP2-TPI | Qwen2-Audio · SlideSpeech L95 | WER (%): 10.76 -> 7.71 | Improved |
| use two-stage pruning (TPI) instead of prompt concatenation (ablation) | SAP2 · SlideSpeech L95, 1 slide | B-WER (%): 7.74 -> 4.71 | Improved |
| use two-stage pruning (TPI) instead of joint pruning-integration (ablation) | SAP2 · SlideSpeech S95, 5 slides | B-WER (%): 9.62 -> 5.18 | Improved |
| use two-stage pruning (TPI) instead of joint pruning-integration (ablation) | SAP2 · SlideSpeech L95 | keyword filtering F1: 20.45 -> 93.85 | Improved |
| extend context from 1 slide to 5 slides (ablation) | SAP2-PC / SAP2-JPI · SlideSpeech | WER: 7.93 -> 7.86 | No clear effect |
| extend context from 1 slide to 5 slides | SAP2-TPI · SlideSpeech L95 | WER (%): 7.92 -> 7.71 | Improved |
| add speech-driven attention pooling (ablation) | Qwen2-Audio PC · SlideSpeech L95, 1 slide | WER (%): 8.22 -> 7.93 | Improved |
| add speech-driven attention pooling (ablation) | Qwen2-Audio TPI · SlideSpeech L95, 5 slides | WER (%): 7.90 -> 7.71 | Improved |
| add speech-driven attention pooling (ablation) | Qwen2-Audio TPI Stage I · SlideSpeech L95 5 slides | training time (s): 51003.8 -> 38267.9 | Improved |
| add speech-driven attention pooling (ablation) | Qwen2-Audio TPI Stage I · SlideSpeech 15-slide test | inference time (s): 1960 -> 1564 | Improved |
| increase pooling window size from 2 to 8 (ablation) | SAP2 · SlideSpeech L95 (5 slides) | WER | Hurt |
| replace CTC-assisted LLM ASR with SAP2-TPI | LibriSpeech · test-other N=100 | B-WER (%): 8.02 -> 6.55 | Improved |
Closest papers by meaning
- Speech Recognition Meets Large Language Model: Benchmarking, Models, and Exploration (AAAI 2025 · similarity 0.86)
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (NeurIPS 2025 · similarity 0.85)
- Unified Speech Recognition: A Single Model for Auditory, Visual, and Audiovisual Inputs (NeurIPS 2024 · similarity 0.85)
- Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference (AAAI 2025 · similarity 0.85)
- Compressed Context Memory for Online Language Model Interaction (ICLR 2024 · similarity 0.85)
- UniCATS: A Unified Context-Aware Text-to-Speech Framework with Contextual VQ-Diffusion and Vocoding (AAAI 2024 · 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.
