ears.py

import os
import tarfile
import datasets
import soundfile as sf  # or any other library that can load audio files

class EarsDataset(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description="EARS dataset containing audio files categorized by speaker IDs.",
            features=datasets.Features({
                'audio': datasets.Audio(sampling_rate=None),  # Use datasets.Audio for audio features
                'spk_id': datasets.Value('string'),  # Include this for training/validation splits
            }),
            supervised_keys=None,
            homepage="https://huggingface.co/datasets/shannan27/ears_dataset",
            citation="Your citation here",
        )

    def _split_generators(self, dl_manager: datasets.DownloadManager):
        # Update URLs to match the paths of the files in your repository
        train_url = "https://huggingface.co/datasets/shannan27/ears_dataset/resolve/main/data/processed_train.tar.gz"
        test_url = "https://huggingface.co/datasets/shannan27/ears_dataset/resolve/main/data/test.tar.gz"
        
        downloaded_train_file = dl_manager.download(train_url)
        downloaded_test_file = dl_manager.download(test_url)
        
        extracted_train_path = os.path.join(dl_manager.manual_dir, 'extracted_train')
        extracted_test_path = os.path.join(dl_manager.manual_dir, 'extracted_test')

        # Extract tar.gz files
        self._extract_archive(downloaded_train_file, extracted_train_path)
        self._extract_archive(downloaded_test_file, extracted_test_path)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"folder_paths": [os.path.join(extracted_train_path, 'train')]},
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"folder_paths": [os.path.join(extracted_test_path, 'blind_testset')]},
            ),
        ]

    def _extract_archive(self, archive_path, extract_to):
        with tarfile.open(archive_path, "r:gz") as tar:
            tar.extractall(path=extract_to)

    def _generate_examples(self, folder_paths):
        """
        Yields examples from the dataset.
        """
        print(folder_paths)
        for folder_path in folder_paths:
            if "train" in folder_path.lower():
                # Processing for train set
                for spk_id in os.listdir(folder_path):
                    spk_folder = os.path.join(folder_path, spk_id)
                    if os.path.isdir(spk_folder):
                        for audio_file in os.listdir(spk_folder):
                            if audio_file.endswith(('.wav', '.mp3')):  # Add other extensions if needed
                                audio_path = os.path.join(spk_folder, audio_file)
                                audio_array, sampling_rate = sf.read(audio_path)  # Load audio file
                                yield f"{spk_id}_{audio_file}", {
                                    'audio': {'array': audio_array, 'sampling_rate': sampling_rate},
                                    'spk_id': spk_id,  # Include spk_id for train set
                                }