seedtts_testset.py

import os
import json
import datasets

_DEFAULT_DATA_DIR = os.path.dirname(__file__)

# 配置YAML里features对应的datasets Features格式
FEATURES = datasets.Features({
    "audio_output_path": datasets.Value("string"),
    "prompt_text": datasets.Value("string"),
    "prompt_audio": datasets.Value("string"),  # 这里用string，后续可cast成 Audio()
    "text_input": datasets.Value("string"),
    "audio_ground_truth": datasets.Value("string"),
})

# YAML里的configs和splits映射成字典
DATASET_CONFIGS = {
    "en": {
        "splits": {
            "test_wer": "data/en_meta.jsonl",
            "test_sim": "data/en_non_para_reconstruct_meta.jsonl",
        }
    },
    "zh": {
        "splits": {
            "test_wer": "data/zh_meta.jsonl",
            "test_sim": "data/zh_non_para_reconstruct_meta.jsonl",
            "test_wer_hardcase": "data/zh_hardcase.jsonl",
        }
    },
}


class SeedTTSDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(name=config_name, version=datasets.Version("1.0.0"))
        for config_name in DATASET_CONFIGS.keys()
    ]

    DEFAULT_CONFIG_NAME = "en"
    
    def _info(self):
        return datasets.DatasetInfo(
            description="Seed-TTS dataset",
            features=FEATURES,
            supervised_keys=None,
        )

    def _split_generators(self, dl_manager):
        config_name = self.config.name

        # 下载并自动解压 repo 中的 data.tgz
        print('Extracting wav files ...')
        archive_path = dl_manager.download_and_extract(f"data/{config_name}.tgz")
        
        # 设置默认 repo的索引 base 路径
        expected_repo_dir = dl_manager._base_path
        index_data_dir = getattr(self.config, "data_dir", None) or expected_repo_dir
        
        # print('archive_path', archive_path)
        wav_data_dir = archive_path
        # import ipdb; ipdb.set_trace()

        # 自动处理压缩包：下载 & 解压到缓存
        wav_data_dir = archive_path
        
        # 提示解压路径
        if dl_manager.manual_dir is not None:
            print(f"[INFO] Using manual_dir: {dl_manager.manual_dir}")
        else:
            print(f"[INFO] Extracted split {config_name} archive to: {archive_path}")
            print("[TIP] To control this cache path, use `load_dataset(..., cache_dir=...)`")

        if config_name not in DATASET_CONFIGS:
            raise ValueError(f"Unknown config name: {config_name}")

        splits = []
        for split_name, relative_path in DATASET_CONFIGS[config_name]["splits"].items():
            file_path = os.path.join(index_data_dir, relative_path)
            if not os.path.isfile(file_path):
                raise FileNotFoundError(
                    f"[ERROR] Missing split file for '{split_name}': expected at '{file_path}'.\n"
                    f"Please make sure all required files exist in the archive."
                )

            splits.append(
                datasets.SplitGenerator(
                    name=split_name,
                    gen_kwargs={"filepath": file_path, "wav_data_dir": wav_data_dir}
                )
            )

        return splits


    def _generate_examples(self, filepath, wav_data_dir):
        with open(filepath, encoding="utf-8") as f:
            for idx, line in enumerate(f):
                item = json.loads(line.strip())

                # 转换相对路径为绝对路径，方便后续cast为Audio
                for key in ["prompt_audio", "audio_ground_truth"]:
                    if not item[key]:
                        continue
                    if key in item and item[key] and not os.path.isabs(item[key]):
                        item[key] = os.path.abspath(os.path.join(wav_data_dir, item[key]))

                yield idx, item