yangfandu commited on
Commit
a5a4d75
·
verified ·
1 Parent(s): fa030bf

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +32 -3
README.md CHANGED
@@ -1,3 +1,32 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Attention2Probability: Attention-Driven Terminology Probability Estimation for Robust Speech-to-Text System
2
+
3
+ <p align="center">
4
+ <a href="https://arxiv.org/abs/2508.18701" alt="paper"><img src="https://img.shields.io/badge/Paper-A2P-blue?logo=arxiv&logoColor=white"/></a>
5
+ <a href="https://huggingface.co/ByteDance/Attention2Probability" alt="Model"><img src="https://img.shields.io/badge/Model-A2P-yellow?logo=huggingface"/></a>
6
+ <a href="https://huggingface.co/datasets/ByteDance/Attention2Probability" alt="Dataset"><img src="https://img.shields.io/badge/Dataset-A2P-yellow?logo=huggingface"/></a>
7
+
8
+
9
+ Attention2Probability (A2P) is a lightweight intervention scheme for speech terminology. The core approach is to use the cross-attention mechanism to retrieve the terms that may appear in the audio and add these terms to the prompt of the llm to complete the term intervention.
10
+
11
+ ## Data description
12
+ This project does not provide audio data for librispeech and aishell2. Please download them from other addresses. All the training data is provided in the data_json folder. The prefix path needs to be modified before use.
13
+
14
+ ## Training step
15
+ For English, the LibriSpeech dataset should first be utilized for pre-training. Subsequently, the second-stage training on LibriSpeech can be conducted by modifying the settings in the dataset configuration.
16
+
17
+ For Chinese, retrieving a single character in isolation lacks practical significance; thus, the Retriever can be directly trained using the Aishell-2 dataset. Finally, the models for both languages are fine-tuned on real-world data.
18
+
19
+
20
+ ## Citation
21
+ If you find A2P useful, please cite the paper:
22
+ ```
23
+ @misc{du2025attention2probabilityattentiondriventerminologyprobability,
24
+ title={Attention2Probability: Attention-Driven Terminology Probability Estimation for Robust Speech-to-Text System},
25
+ author={Yanfan Du and Jun Zhang and Bin Wang and Jin Qiu and Lu Huang and Yuan Ge and Xiaoqian Liu and Tong Xiao and Jingbo Zhu},
26
+ year={2025},
27
+ eprint={2508.18701},
28
+ archivePrefix={arXiv},
29
+ primaryClass={cs.CL},
30
+ url={https://arxiv.org/abs/2508.18701},
31
+ }
32
+ ```