Datasets:

Hezep
/

AudioMarathon

@@ -1,338 +0,0 @@
-"""
-AudioMarathon Dataset Loading Script for Hugging Face
-支持10个音频理解任务的数据集加载
-"""
-import datasets
-import json
-import os
-import glob
-from pathlib import Path
-_CITATION = """
-@misc{audio_marathon_2025,
-  title={AudioMarathon: A Comprehensive Audio Understanding Benchmark},
-  author={Your Team},
-  year={2025}
-}
-"""
-_DESCRIPTION = """
-AudioMarathon is a comprehensive multi-task audio understanding benchmark containing 10 tasks:
-1. **RACE**: Reading Comprehension QA from audio
-2. **HAD**: Half-truth Audio Detection (authenticity classification)
-3. **GTZAN**: Music Genre Classification
-4. **TAU**: Acoustic Scene Classification
-5. **VESUS**: Emotion Recognition from speech
-6. **SLUE**: Named Entity Recognition from speech
-7. **DESED**: Sound Event Detection
-8. **VoxCeleb-Gender**: Speaker Gender Identification
-9. **VoxCeleb-Age**: Speaker Age Classification
-10. **LibriSpeech**: Automatic Speech Recognition (ASR)
-Total: ~6,000 audio samples, 60GB data
-"""
-_HOMEPAGE = "https://github.com/your-repo/audio-marathon"
-_LICENSE = "CC-BY-4.0"
-# 数据文件路径映射
-_URLS = {
-    "race": "Dataset/race_audio/race_benchmark.json",
-    "had": "Dataset/HAD/concatenated_audio/had_audio_classification_task.json",
-    "gtzan": "Dataset/GTZAN/concatenated_audio/music_genre_classification_meta.json",
-    "tau": "Dataset/TAU/acoustic_scene_task_meta.json",
-    "vesus": "Dataset/VESUS/audio_emotion_dataset.json",
-    "slue": "Dataset/SLUE/merged_audio_data.json",
-    "desed": "Dataset/DESED/DESED_dataset/concatenated_audio/desed_sound_event_detection_task.json",
-    "voxceleb_gender": "Dataset/VoxCeleb/concatenated_audio/gender_id_task_meta.json",
-    "voxceleb_age": "Dataset/VoxCeleb/concatenated_audio_age/age_classification_task_meta.json",
-    "librispeech": "Dataset/librispeech-long",  # 目录而非JSON
-}
-class AudioMarathonConfig(datasets.BuilderConfig):
-    """BuilderConfig for AudioMarathon."""
-    def __init__(self, **kwargs):
-        super(AudioMarathonConfig, self).__init__(**kwargs)
-class AudioMarathon(datasets.GeneratorBasedBuilder):
-    """AudioMarathon Dataset - 10个音频理解任务"""
-    VERSION = datasets.Version("1.0.0")
-    BUILDER_CONFIGS = [
-        AudioMarathonConfig(name="race", version=VERSION,
-                            description="Reading comprehension from audio"),
-        AudioMarathonConfig(name="had", version=VERSION,
-                            description="Half-truth audio detection"),
-        AudioMarathonConfig(name="gtzan", version=VERSION,
-                            description="Music genre classification"),
-        AudioMarathonConfig(name="tau", version=VERSION,
-                            description="Acoustic scene classification"),
-        AudioMarathonConfig(name="vesus", version=VERSION,
-                            description="Emotion recognition"),
-        AudioMarathonConfig(name="slue", version=VERSION,
-                            description="Named entity recognition"),
-        AudioMarathonConfig(name="desed", version=VERSION,
-                            description="Sound event detection"),
-        AudioMarathonConfig(name="voxceleb_gender", version=VERSION,
-                            description="Speaker gender identification"),
-        AudioMarathonConfig(name="voxceleb_age", version=VERSION,
-                            description="Speaker age classification"),
-        AudioMarathonConfig(name="librispeech", version=VERSION,
-                            description="Automatic speech recognition"),
-        AudioMarathonConfig(name="all", version=VERSION,
-                            description="All 10 tasks combined"),
-    ]
-    DEFAULT_CONFIG_NAME = "all"
-    def _info(self):
-        """定义数据集特征"""
-        features = datasets.Features({
-            "audio": datasets.Audio(sampling_rate=16000),
-            "question": datasets.Value("string"),
-            "options": datasets.Sequence(datasets.Value("string")),
-            "answer": datasets.Value("string"),
-            "task_name": datasets.Value("string"),
-            "dataset_name": datasets.Value("string"),
-            "uniq_id": datasets.Value("int32"),
-            "metadata": datasets.Value("string"),  # JSON string for flexible metadata
-        })
-        return datasets.DatasetInfo(
-            description=_DESCRIPTION,
-            features=features,
-            homepage=_HOMEPAGE,
-            license=_LICENSE,
-            citation=_CITATION,
-        )
-    def _split_generators(self, dl_manager):
-        """下载并组织数据分片 - 只返回test分片"""
-        config_name = self.config.name
-        # 根据配置选择要加载的任务
-        if config_name == "all":
-            tasks_to_load = list(_URLS.keys())
-        else:
-            tasks_to_load = [config_name]
-        # 准备数据文件路径（不使用dl_manager.download，直接使用本地路径）
-        data_paths = {}
-        for task in tasks_to_load:
-            if task in _URLS:
-                # 如果是相对路径，转换为绝对路径
-                path = _URLS[task]
-                if not os.path.isabs(path):
-                    # 假设脚本在AudioMarathon目录下
-                    base_dir = os.path.dirname(os.path.abspath(__file__))
-                    path = os.path.join(base_dir, path)
-                data_paths[task] = path
-        # 只返回test分片，不返回train和validation
-        return [
-            datasets.SplitGenerator(
-                name=datasets.Split.TEST,
-                gen_kwargs={
-                    "data_paths": data_paths,
-                },
-            ),
-        ]
-    def _generate_examples(self, data_paths):
-        """生成数据样本"""
-        idx = 0
-        for task_name, data_path in data_paths.items():
-            print(f"Loading task: {task_name} from {data_path}")
-            if task_name == "librispeech":
-                # LibriSpeech特殊处理：从目录结构加载
-                for example in self._load_librispeech(data_path):
-                    yield idx, example
-                    idx += 1
-            else:
-                # 其他任务：从JSON加载
-                for example in self._load_from_json(task_name, data_path):
-                    yield idx, example
-                    idx += 1
-    def _load_from_json(self, task_name, json_path):
-        """从JSON文件加载数据"""
-        if not os.path.exists(json_path):
-            print(f"Warning: JSON file not found: {json_path}")
-            return
-        try:
-            with open(json_path, 'r', encoding='utf-8') as f:
-                data = json.load(f)
-        except Exception as e:
-            print(f"Error loading {json_path}: {e}")
-            return
-        # 获取基础目录（用于构建音频文件路径）
-        base_dir = os.path.dirname(json_path)
-        # 根据数据格式提取样本列表
-        if isinstance(data, list):
-            samples = data
-        elif isinstance(data, dict):
-            # DESED的特殊格式
-            if "tasks" in data:
-                samples = data["tasks"]
-            # HAD的特殊格式
-            elif "samples" in data:
-                samples = data["samples"]
-            else:
-                print(f"Warning: Unknown JSON format for {task_name}")
-                return
-        else:
-            print(f"Warning: Unexpected data type for {task_name}")
-            return
-        print(f"  Found {len(samples)} samples in {task_name}")
-        # 处理每个样本
-        for sample in samples:
-            # 构建音频完整路径
-            audio_rel_path = sample.get("path", "")
-            if not audio_rel_path:
-                continue
-            audio_path = os.path.join(base_dir, audio_rel_path)
-            # 跳过不存在的文件
-            if not os.path.exists(audio_path):
-                print(f"Warning: Audio file not found: {audio_path}")
-                continue
-            # 提取options（支持多种格式）
-            options = []
-            if "options" in sample:
-                options = sample["options"]
-            else:
-                # 从choice_a, choice_b等构建
-                for choice in ["choice_a", "choice_b", "choice_c", "choice_d", "choice_e"]:
-                    if choice in sample:
-                        options.append(sample[choice])
-            # 提取答案
-            answer = sample.get("answer", sample.get("answer_gt", ""))
-            # 构建元数据
-            metadata = {
-                k: v for k, v in sample.items()
-                if k not in ["path", "question", "options", "answer", "answer_gt",
-                            "choice_a", "choice_b", "choice_c", "choice_d", "choice_e"]
-            }
-            yield {
-                "audio": audio_path,
-                "question": sample.get("question", ""),
-                "options": options,
-                "answer": answer,
-                "task_name": sample.get("task_name", task_name),
-                "dataset_name": sample.get("dataset_name", task_name.upper()),
-                "uniq_id": sample.get("uniq_id", -1),
-                "metadata": json.dumps(metadata, ensure_ascii=False)
-            }
-    def _load_librispeech(self, data_dir):
-        """加载LibriSpeech数据"""
-        if not os.path.exists(data_dir):
-            print(f"Warning: LibriSpeech directory not found: {data_dir}")
-            return
-        # 可能的LibriSpeech子目录
-        possible_subdirs = [
-            "test-clean",
-            "test-other",
-            "dev-clean",
-            "dev-other",
-        ]
-        # 查找说话人目录
-        speaker_dirs = []
-        # 检查是否直接包含说话人文件夹（数字命名）
-        direct_speakers = sorted([
-            d for d in glob.glob(os.path.join(data_dir, "*"))
-            if os.path.isdir(d) and os.path.basename(d).isdigit()
-        ])
-        if direct_speakers:
-            speaker_dirs = direct_speakers
-        else:
-            # 检查子目录
-            for subdir in possible_subdirs:
-                subdir_path = os.path.join(data_dir, subdir)
-                if os.path.exists(subdir_path):
-                    speakers = sorted([
-                        d for d in glob.glob(os.path.join(subdir_path, "*"))
-                        if os.path.isdir(d) and os.path.basename(d).isdigit()
-                    ])
-                    speaker_dirs.extend(speakers)
-        if not speaker_dirs:
-            print(f"Warning: No speaker directories found in {data_dir}")
-            return
-        print(f"  Found {len(speaker_dirs)} speakers in LibriSpeech")
-        # 处理每个说话人目录
-        sample_idx = 0
-        for speaker_dir in speaker_dirs:
-            speaker_id = os.path.basename(speaker_dir)
-            # 查找章节目录
-            chapter_dirs = sorted([
-                d for d in glob.glob(os.path.join(speaker_dir, "*"))
-                if os.path.isdir(d)
-            ])
-            for chapter_dir in chapter_dirs:
-                chapter_id = os.path.basename(chapter_dir)
-                # 查找转录文件
-                trans_file = os.path.join(chapter_dir, f"{speaker_id}-{chapter_id}.trans.txt")
-                if not os.path.exists(trans_file):
-                    continue
-                # 读取转录文本
-                transcripts = {}
-                try:
-                    with open(trans_file, 'r', encoding='utf-8') as f:
-                        for line in f:
-                            parts = line.strip().split(' ', 1)
-                            if len(parts) == 2:
-                                utt_id, text = parts
-                                transcripts[utt_id] = text
-                except Exception as e:
-                    print(f"Error reading {trans_file}: {e}")
-                    continue
-                # 查找对应的音频文件
-                for utt_id, transcript in transcripts.items():
-                    audio_file = os.path.join(chapter_dir, f"{utt_id}.flac")
-                    if os.path.exists(audio_file):
-                        yield {
-                            "audio": audio_file,
-                            "question": "Transcribe this audio accurately. Remove hesitation words like 'um', 'uh'.",
-                            "options": [],  # ASR任务通常没有选项
-                            "answer": transcript,
-                            "task_name": "Automatic_Speech_Recognition",
-                            "dataset_name": "LibriSpeech",
-                            "uniq_id": sample_idx,
-                            "metadata": json.dumps({
-                                "speaker_id": speaker_id,
-                                "chapter_id": chapter_id,
-                                "utterance_id": utt_id
-                            })
-                        }
-                        sample_idx += 1