falAR.py

import datasets

DESCRIPTION = """
FalAR dataset with a unified training split.
All original train_* splits are merged on the fly via streaming.
"""

class FalARDataset(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(name="default", version=datasets.Version("1.0.0")),
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=DESCRIPTION,
            features={
                "id": datasets.Value("string"),
                "text": datasets.Value("string"),
                "speaker": datasets.Value("string"),
                "wav": datasets.Audio(sampling_rate=16000),
            },
        )

    def _split_generators(self, dl_manager):

        # Automatically generated list of training splits
        train_splits = [
            "train_0", "train_1", "train_10", "train_11", "train_12", "train_13", "train_14", "train_15", "train_2", "train_3", "train_4", "train_5", "train_6", "train_7", "train_8", "train_9"
        ]

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"splits": train_splits},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"splits": ["dev"]},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"splits": ["test"]},
            ),
        ]

    def _generate_examples(self, splits):

        # Load each split using streaming = avoids downloading huge data
        loaded = [
            datasets.load_dataset("inesc-id/FalAR", split=split, streaming=True)
            for split in splits
        ]

        idx = 0
        for ds in loaded:
            for example in ds:
                yield idx, example
                idx += 1