first commit

.gitattributes

@@ -57,3 +57,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,129 @@
+# SAP²-ASR Dataset
+
+本数据集用于 **SAP² (Speech-Aware Long Context Pruning and Integration)** 方法的上下文感知自动语音识别（ASR）研究。
+
+## 📖 简介
+
+SAP² 是一个用于上下文感知自动语音识别的新框架，能够动态剪枝并集成相关的上下文关键词。该方法解决了在特定领域场景（如会议演讲）中利用长上下文信息的挑战，这些场景中大量来自OCR的文本上下文既包含相关信息，也包含大量噪声。
+
+### 核心特性
+
+- **语音感知上下文剪枝**：动态过滤来自OCR的文本上下文，仅保留与语音内容直接相关的关键词
+- **跨模态上下文压缩**：使用语音驱动注意力池化（Speech-Driven Attention-based Pooling）将大量文本输入压缩为简洁的、与语音相关的上下文嵌入
+- **最先进的性能**：在 SlideSpeech 数据集上达到 7.71% 的词错误率（WER），在 LibriSpeech 数据集上达到 1.12% 的 WER，相比非上下文基线，在偏向关键词识别方面相对提升了 41.1%
+
+## 📊 数据集结构
+
+本数据集包含两个主要子数据集：
+
+### SlideSpeech
+
+- **来源**：SlideSpeech 是一个包含幻灯片的大规模音视频语料库，包含 1,705 个视频，超过 1,000 小时的音频，其中包括 473 小时的高质量转录语音
+- **数据格式**：JSON 格式，包含音频路径和带上下文关键词的对话格式
+- **目录结构**：
+  - `slidespeech_L95/`: 原始数据（对应论文中PC）
+  - `slidespeech_L95_filter/`: 训练TPI第一阶段
+  - `slidespeech_L95_filtered_train`: 训练TPI第二阶段
+  - `slidespeech_L95_5slides/`: 5张幻灯片版本
+  - `slidespeech_L95_multitask/`: 多任务版本（对应论文中的JPI）
+
+### LibriSpeech
+
+- **来源**：LibriSpeech 是一个大规模英语朗读语音语料库，源自 LibriVox 项目的有声读物
+- **数据格式**：JSON 格式，包含训练、验证和测试集的不同配置
+- **目录结构**：
+  - `train-clean-460_*.json`: 训练集（clean，460小时）
+  - `train-other-500_*.json`: 训练集（other，500小时）
+  - `dev-clean_*.json`, `dev-other_*.json`: 验证集
+  - `test-clean_*.json`, `test-other_*.json`: 测试集（不同规模：100, 500, 1000, 2000条）
+
+### 数据格式示例
+
+```json
+{
+  "messages": [
+    {
+      "role": "user",
+      "content": "<audio>/path/to/audio.wav</audio>Transcribe speech to text according to keywords may appear in the utterance. Possible keywords are: <|startofcontext|>keyword1 keyword2 keyword3<|endofcontext|>"
+    },
+    {
+      "role": "assistant",
+      "content": "transcribed text"
+    }
+  ],
+  "audios": "/path/to/audio.wav"
+}
+```
+
+**关键标记**：
+- `<|startofcontext|>` 和 `<|endofcontext|>`: 用于标记上下文关键词的特殊标记
+- `<audio>...</audio>`: 音频文件路径标记
+
+## 🚀 使用方法
+
+### 加载数据集
+
+```python
+import json
+
+# 加载 SlideSpeech 数据集
+with open('slidespeech/slidespeech_L95_filter/train.json', 'r') as f:
+    slidespeech_train = json.load(f)
+
+# 加载 LibriSpeech 数据集
+with open('librispeech/train-clean-460_filter.json', 'r') as f:
+    librispeech_train = json.load(f)
+```
+
+### 与 SAP² 模型一起使用
+
+详细的使用说明、训练和推理代码请参考：
+
+**🔗 [SAP²-ASR GitHub Repository](https://github.com/jymh/SAP2-ASR.git)**
+
+该仓库包含：
+- 完整的模型实现代码
+- 训练和推理脚本
+- 数据预处理工具
+- 评估脚本
+- 详细的使用文档
+
+## 📎 引用
+
+如果您在研究中使用了本数据集，请引用以下论文：
+
+```bibtex
+@article{rong2025speechaware,
+  title={Speech-Aware Long Context Pruning and Integration for Contextualized Automatic Speech Recognition},
+  author={Rong, Yiming and Zhang, Yixin and Wang, Ziyi and Jiang, Deyang and Zhao, Yunlong and Wu, Haoran and Zhou, Shiyu and Xu, Bo},
+  journal={arXiv preprint arXiv:2511.11139},
+  year={2025}
+}
+```
+
+**论文链接**：[https://www.arxiv.org/abs/2511.11139](https://www.arxiv.org/abs/2511.11139)
+
+## 📚 相关资源
+
+- **代码仓库**：[https://github.com/jymh/SAP2-ASR.git](https://github.com/jymh/SAP2-ASR.git)
+- **论文**：[arXiv:2511.11139](https://www.arxiv.org/abs/2511.11139)
+- **SlideSpeech 原始数据集**：[https://slidespeech.github.io/](https://slidespeech.github.io/)
+- **LibriSpeech 原始数据集**：[OpenSLR](https://www.openslr.org/12/)
+
+## 🏛 许可证
+
+本数据集的使用请遵循原始数据集的许可证要求。对于 SlideSpeech 和 LibriSpeech，请参考其原始资源页面的许可证说明。
+
+## ⚠️ 注意事项
+
+1. 音频文件路径可能需要根据实际环境进行调整
+2. 数据集文件较大，请确保有足够的存储空间
+3. 使用前请仔细阅读 [GitHub 仓库](https://github.com/jymh/SAP2-ASR.git) 中的详细文档
+
+---
+
+**更多信息和使用示例，请访问 [SAP²-ASR GitHub Repository](https://github.com/jymh/SAP2-ASR.git)**
+
 ---
 license: apache-2.0
 ---

librispeech/dev-clean_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:494079e66f5983e481a2423858292e68b7a271e9ba1a704182a0ab3c078860f0
+size 16995084

librispeech/dev-clean_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:214bd3858a522b3411d9ba2a6a8e3492e155239bbb637a74ebc1936a63893e83
+size 1361688

librispeech/dev-clean_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:569c4c5bc96b7a8bc1b3edcc51d7114f38c481e7d1e4d4bc6507e08e4c28af51
+size 17284317

librispeech/dev-other_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:047450d5d1742788e9600fd1430378a24499c9f9ad51a953f49d1dd1d60c6b46
+size 17947988

librispeech/dev-other_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67bc165d5ab132c65ae18d9d8ebbe6a5496fb4a3d4084e86bb49a756b80db013
+size 1378646

librispeech/dev-other_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf629643a4d0b12969776917ddc013248c17025d170366c19bdfb0b7f34ac388
+size 18225434

librispeech/test-clean_1000_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dc449ac48a83d8a5db2f435932416b7d99e198114571a249ec9c6986b304239
+size 26661553

librispeech/test-clean_1000_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:868eb9e72ca917406659dd93b1fda09acec7b4f063a8e9c711547f8b7680b215
+size 1322871

librispeech/test-clean_1000_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e056ed82f802a8cf61e6714de09caef1a78b68eda6349e6447e7a86aa1849de8
+size 26941872

librispeech/test-clean_100_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aefeed8018ffe5fb6f8b40803537524e0bbe1bf444e8f93ecc2dbb9524c71a28
+size 3649107

librispeech/test-clean_100_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:868eb9e72ca917406659dd93b1fda09acec7b4f063a8e9c711547f8b7680b215
+size 1322871

librispeech/test-clean_100_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:066d66baf4f5f632f377147ad7c9e9c3c831202b4de1f0f2ff4ce18e1b156859
+size 3929426

librispeech/test-clean_2000_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ae13a50bd8a21810aa3a5eda7dec09319b87c53b735777c3a5a395fb8e21861
+size 52137181

librispeech/test-clean_2000_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:868eb9e72ca917406659dd93b1fda09acec7b4f063a8e9c711547f8b7680b215
+size 1322871

librispeech/test-clean_2000_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a405223fa0156c25ab722ab10a08d11768c1b103192568ea838a031f67741064
+size 52417500

librispeech/test-clean_500_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05bf5c581551335dcf0f1dfd4eff64f7ddc73c128f9ef9340cf27c78c4fae587
+size 13884322

librispeech/test-clean_500_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:868eb9e72ca917406659dd93b1fda09acec7b4f063a8e9c711547f8b7680b215
+size 1322871

librispeech/test-clean_500_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbfaf6cfaec254deecc8e25970430d7a7a070dff5ea57af05fc9aabdd738a59e
+size 14164641

librispeech/test-other_1000_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f2fb2bd24bcf99f9fdfd6f43027055df124c2a03299b57493a1f3c3cb6bc5ee
+size 29880756

librispeech/test-other_1000_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186f155270c515fabcfe2a4e2204ac0e80964a000e6de80a39fa2e6cc9e959b3
+size 1424509

librispeech/test-other_1000_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d8796dc23c6cc4b154a265ffbde3061480eaf4ec74f2063852542344a4767edc
+size 30164633

librispeech/test-other_100_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4719130b77ddbcf44732cf21a536430007d4cfebbb4b484742d5409a639b2f3
+size 4071171

librispeech/test-other_100_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186f155270c515fabcfe2a4e2204ac0e80964a000e6de80a39fa2e6cc9e959b3
+size 1424509

librispeech/test-other_100_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c27aac655cfd574cad309e26cd9f155a1bbb6b01e64cc2d950b9a2cf37d76a5
+size 4355048

librispeech/test-other_2000_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72dfb730e0167d794205ebe98077d39cf3969616a819021969192df7a68b5e73
+size 58468062

librispeech/test-other_2000_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186f155270c515fabcfe2a4e2204ac0e80964a000e6de80a39fa2e6cc9e959b3
+size 1424509

librispeech/test-other_2000_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40c8545d72c7c00a7b61f1a5509910116113ef62d488468fdc0baf8028ed0a40
+size 58751939

librispeech/test-other_500_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd7547955be409c296402b08666163d2fa32749afc5db9b6e77fc0054d3e251a
+size 15552466

librispeech/test-other_500_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186f155270c515fabcfe2a4e2204ac0e80964a000e6de80a39fa2e6cc9e959b3
+size 1424509

librispeech/test-other_500_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bf0b30c6c20a654e6702a8ce96865b174dcb9b6f79c5bab266b7ef41df87644
+size 15836343

librispeech/train-clean-460_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18f739fa3b818a717d89ae028b4cf13c40ad202e257a4e8e414452ea2bf490d5
+size 836153907

librispeech/train-clean-460_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2582b8382240c4c280343b6e87f8b8046935aa3ad5a3a6295a9d865c4d8a2914
+size 81066086

librispeech/train-clean-460_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b535f5e4aedd1a7697c42624e1c4a236796d02caa6c695f29d1e1ff41e584342
+size 858998364

librispeech/train-other-500_filter.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:882a52fa224a9f38c51cdc130877d95a126236e91e22983c8f41e3af87a1a244
+size 936464758

librispeech/train-other-500_filtered_asr.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed6d94d9cbb825af3ea86253d0b2c2eda618e1444bb9eda411c5954905fd7cd9
+size 88132611

librispeech/train-other-500_pc.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ddd785c26d91699815157d45b94ba7f4b36d6d49ffb5b57d23286a133004197
+size 960631559

slidespeech/slidespeech_L95/dev.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a0b71a9c7348bc9c497f933f1b5bd53f8e96231c7225be635c5462e5fb71815
+size 1401937

slidespeech/slidespeech_L95/test.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf6ebe808b4877b618cb1c5fd2bd87bb36eec0f0525df80bd8ae234cbd7d558c
+size 2487403

slidespeech/slidespeech_L95/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e35ec7372655356a14a2f6ab2b7ab7412169ce8abb1982014c86c371ec6e844b
+size 307037255

slidespeech/slidespeech_L95_5slides/dev.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af2543de7cee371702862f96a143d5206c9fd76aa87e355d9c03ad5c2a560733
+size 2928654

slidespeech/slidespeech_L95_5slides/test.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3743038e77227a6a43bd98d14a83e334245094692f9cb7696c7def7a3c28002
+size 5615310

slidespeech/slidespeech_L95_5slides/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d9434e2c49116ef76628ae73697b90ef922c45eef44f79cb82162930609ec55
+size 708440282

slidespeech/slidespeech_L95_5slides_filter/S95.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bff6bb7a060dcd16af3a574ecbe4814f685efec45c6ef2620acc4da2212fb038
+size 145470666

slidespeech/slidespeech_L95_5slides_filter/dev.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c970c517d93cb9922fd6ad117128d43367af2a46f4d05c9da9851717bb72bf7
+size 2676197

slidespeech/slidespeech_L95_5slides_filter/test.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a77df39f0da60abae746f8dcfb559ffbcb80af4f8638116600420738208398e
+size 5171425

slidespeech/slidespeech_L95_5slides_filter/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a8428d6b11b40b4f274008cb89129d1bf65e05fd927ad29191e40b17a5302f9
+size 676787747

slidespeech/slidespeech_L95_5slides_filtered_train/S95.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8b9f6435adc29c1a410d580432b9712309600a343861e79326d069045e584e3
+size 41677744

slidespeech/slidespeech_L95_5slides_filtered_train/count_keywords_after_filter_251011.py

@@ -0,0 +1,146 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+统计JSON文件中每条样本的<|startofcontext|><|endofcontext|>间的平均词数
+"""
+
+import json
+import re
+from typing import List, Dict, Any
+
+
+def count_words_in_context(text: str) -> int:
+    """
+    统计文本中<|startofcontext|>和<|endofcontext|>之间的词数
+    
+    Args:
+        text: 输入文本
+        
+    Returns:
+        context区域内的词数
+    """
+    # 使用正则表达式匹配<|startofcontext|>和<|endofcontext|>之间的内容
+    pattern = r'<\|startofcontext\|>(.*?)<\|endofcontext\|>'
+    matches = re.findall(pattern, text, re.DOTALL)
+    
+    total_words = 0
+    for match in matches:
+        # 去除首尾空白字符，然后按空白字符分割计算词数
+        words = match.strip().split(", ")
+        
+        words = [] if words == [''] else words
+        print(words)
+        total_words += len(words)
+    
+    return total_words
+
+
+def analyze_json_file(file_path: str) -> Dict[str, Any]:
+    """
+    分析JSON文件中的context词数统计
+    
+    Args:
+        file_path: JSON文件路径
+        
+    Returns:
+        包含统计结果的字典
+    """
+    print(f"正在读取文件: {file_path}")
+    
+    with open(file_path, 'r', encoding='utf-8') as f:
+        data = json.load(f)
+    
+    print(f"总样本数: {len(data)}")
+    
+    word_counts = []
+    samples_with_context = 0
+    samples_without_context = 0
+    
+    for i, sample in enumerate(data):
+        if i % 1000 == 0:
+            print(f"处理进度: {i}/{len(data)}")
+        
+        # 检查每条消息中的content
+        sample_word_count = 0
+        has_context = False
+        
+        for message in sample.get('messages', []):
+            content = message.get('content', '')
+            word_count = count_words_in_context(content)
+            sample_word_count += word_count
+            
+            if '<|startofcontext|>' in content and '<|endofcontext|>' in content:
+                has_context = True
+        
+        if has_context:
+            samples_with_context += 1
+            word_counts.append(sample_word_count)
+        else:
+            samples_without_context += 1
+    
+    # 计算统计信息
+    if word_counts:
+        avg_words = sum(word_counts) / len(word_counts)
+        max_words = max(word_counts)
+        min_words = min(word_counts)
+        
+        # 计算中位数
+        sorted_counts = sorted(word_counts)
+        n = len(sorted_counts)
+        median_words = (sorted_counts[n//2 - 1] + sorted_counts[n//2]) / 2 if n % 2 == 0 else sorted_counts[n//2]
+    else:
+        avg_words = max_words = min_words = median_words = 0
+    
+    results = {
+        'total_samples': len(data),
+        'samples_with_context': samples_with_context,
+        'samples_without_context': samples_without_context,
+        'average_words_per_sample': avg_words,
+        'max_words_per_sample': max_words,
+        'min_words_per_sample': min_words,
+        'median_words_per_sample': median_words,
+        'word_counts_distribution': word_counts[:10] if len(word_counts) > 10 else word_counts  # 显示前10个样本的词数
+    }
+    
+    return results
+
+
+def main():
+    """主函数"""
+    file_path = "/data8/rym/Projects/ms-swift/data/add_context_token/slidespeech_L95_5slides_filtered_train/test_from_window_size2_no_param.json"
+    
+    try:
+        results = analyze_json_file(file_path)
+        
+        print("\n" + "="*60)
+        print("统计结果:")
+        print("="*60)
+        print(f"总样本数: {results['total_samples']}")
+        print(f"包含context的样本数: {results['samples_with_context']}")
+        print(f"不包含context的样本数: {results['samples_without_context']}")
+        print(f"平均每条样本context区域词数: {results['average_words_per_sample']:.2f}")
+        print(f"最大词数: {results['max_words_per_sample']}")
+        print(f"最小词数: {results['min_words_per_sample']}")
+        print(f"中位数词数: {results['median_words_per_sample']:.2f}")
+        
+        if results['word_counts_distribution']:
+            print(f"\n前10个样本的context词数分布:")
+            for i, count in enumerate(results['word_counts_distribution']):
+                print(f"  样本 {i+1}: {count} 词")
+        
+        # # 保存详细结果到文件
+        # output_file = file_path.replace('.json', '_context_word_stats.json')
+        # with open(output_file, 'w', encoding='utf-8') as f:
+        #     json.dump(results, f, ensure_ascii=False, indent=2)
+        
+        # print(f"\n详细统计结果已保存到: {output_file}")
+        
+    except Exception as e:
+        print(f"处理文件时出错: {e}")
+        return 1
+    
+    return 0
+
+
+if __name__ == "__main__":
+    exit(main())