第 11 章：附录——速查表、模板与资源 (Reference & Resources)

# Dataset Card for [数据集名称]

## 1. 基本信息 (Metadata)

- **Dataset Name**: `[例如: UrbanSound8K-Clean]`
- **Version**: `[e.g., v1.0.0]`
- **Language/Region**: `[e.g., English (US), Chinese (Mandarin)]`
- **Task Category**: `[e.g., SED, ASR, TTS]`
- **License**: `[e.g., CC-BY-4.0]`
- **Total Duration**: `[e.g., 8.5 hours]`
- **Sample Rate**: `[e.g., 44.1kHz (original), provided as 16kHz]`

## 2. 数据构成与划分 (Structure & Splits)
| Split | Samples | Duration | Speakers/Classes | Description |

| Split | Samples | Duration | Speakers/Classes | Description |
| :--- | :--- | :--- | :--- | :--- |
| `train` | 10,000 | 80h | 50 speakers | 主训练集 |
| `dev` | 1,000 | 8h | 10 speakers | 验证集 (Seen speakers) |
| `test` | 1,000 | 8h | 10 speakers | 测试集 (Unseen speakers) |

## 3. 标注详情 (Annotations)

- **Format**: `[e.g., JSONL, TextGrid]`
- **Granularity**: `[e.g., Strong labels with timestamps / Weak labels]`
- **Label Schema**: `[Link to class mapping file]`
- **Annotation Process**:
  - [ ] Automatic Transcribed
  - [x] Human Verified (Single annotator)
  - [ ] Human Verified (Double blind + Arbitration)

## 4. 采集与处理 (Collection & Processing)

- **Source**: `[e.g., Youtube (Creative Commons filter)]`
- **Recording Environment**: `[e.g., Street, Cafe, Home]`
- **Preprocessing**:
  - VAD applied: Yes/No
  - Noise removed: Yes/No
  - Loudness normalization: -23 LUFS

## 5. 局限性与伦理 (Limitations & Ethics)

- **Bias**: 数据中男性声音占比 80%。
- **Privacy**: 所有 PII (姓名、电话) 已通过自动检测+人工听测掩盖。
- **Known Issues**: 约 5% 的样本背景噪音极大约盖过人声。

# Benchmark Card for [基准名称]

## 1. 评测目标 (Objective)
评估模型在 `[场景，如：强混响环境]` 下的 `[能力，如：指令跟随]` 能力。

## 2. 协议定义 (Protocol)

- **Input Modality**: Audio (WAV) + Text Instruction
- **Output Modality**: Text
- **Shot Setting**:
  - [x] Zero-shot (标准设置)
  - [ ] Few-shot (允许 5-shot，样本需来自 `assets/few_shot_examples.json`)
- **Evaluation Set**: `manifest/test_benchmark_v2.jsonl` (MD5: `...`)

## 3. 推理配置 (Inference Config)

- **Sampling Rate**: 统一重采样至 16kHz
- **Prompt Template**:
  > "Listen to the audio. Question: {question} Answer:"

- **Decoding Strategy**: Greedy search (Temperature=0) for reproducibility.

## 4. 评价指标 (Metrics)
| Metric | Implementation | Scope |

| Metric | Implementation | Scope |
| :--- | :--- | :--- |
| **Accuracy** | Exact Match (case-insensitive) | Main Metric |
| **F1-Score** | `sklearn.metrics.f1_score` (macro) | Analysis |
| **RTF** | Real Time Factor on A100 | Efficiency |

## 5. 基线 (Baselines)
| Model | Accuracy | F1 | Notes |

| Model | Accuracy | F1 | Notes |
| :--- | :--- | :--- | :--- |
| Qwen-Audio | 65.2% | 0.60 | Zero-shot |
| SALMONN | 62.1% | 0.58 | - |

my_audio_project/
├── assets/                  # 存放 prompt 模板、类别映射表 (class_map.json)
├── data/                    # 【只读】原始数据挂载点 (通常是软链接)
│   ├── aishell/
│   └── audioset/
├── manifest/                # 【核心】数据索引 (JSONL/CSV)
│   ├── v1.0/
│   │   ├── train.jsonl      # 格式: {"wav": "path", "txt": "...", "dur": 5.2}
│   │   └── test.jsonl
│   └── processing_scripts/  # 生成 manifest 的脚本
├── recipes/                 # 【核心】训练配方 (YAML + Shell)
│   ├── asr_conformer/
│   │   ├── config.yaml      # 模型超参、学习率
│   │   └── run.sh           # 启动脚本
│   └── tts_vits/
├── checkpoints/             # 模型产出 (按 exp_id 隔离)
│   └── exp_20231027_baseline/
│       ├── tensorboard/
│       ├── best_model.pt
│       └── config_saved.yaml
├── utils/                   # 通用工具
│   ├── audio_io.py          # 统一音频读取 (Torchaudio/Librosa 封装)
│   └── metric_calc.py       # 统一指标计算器
└── README.md                # 项目入口文档

{"id": "utt_001", "audio_filepath": "/data/wavs/001.wav", "text": "Hello world", "duration": 3.5, "speaker_id": "spk_1"}

{"id": "qa_001", "audio": "/data/wavs/rain.wav", "instruction": "描述这段声音", "answer": "这是大雨落在屋顶的声音。", "task": "captioning", "dataset": "clotho"}