mumospee_small.py

import csv
import os
import datasets
import pandas as pd

# Metadata

_DESCRIPTION = """\

Mumospee is a continuously growing, comprehensive, multilingual dataset across different modalities. 
This is the small version include no more 1000 rows.

"""

_LICENSE = "Creative Commons Attribution 4.0 International"

# Defining categories or tags for filtering
_LANGUAGES = ["en", "bg", "de", "es", "it", "ar"]  # Example languages
_TAGS = ["CoVoST", "GigaSpeech", "peoples_speech", "Librispeech", "LibriTTS", "Emilia", "MOSEL"]

_SPLITS = ["train", "validation", "test"]

# BuilderConfig class for your dataset
class MumospeeDatasetConfig(datasets.BuilderConfig):
    def __init__(self, name, *args, **kwargs):
        super().__init__(name=name, *args, **kwargs)
        self.language = kwargs.get("language", None) # Add language as a key in the config
        self.tag = kwargs.get("tag", None)  # Add tag as a key in the config



class MumospeeDataset(datasets.GeneratorBasedBuilder):
    """Your custom dataset for Hugging Face based on the CSV metadata and audio URLs."""

    VERSION = datasets.Version("1.0.0")

    # Define the available configurations (could be subsets like split or language)
    BUILDER_CONFIGS = [
        MumospeeDatasetConfig(
            name="default",
            version=datasets.Version("1.0.0"),
            description=_DESCRIPTION)
        ]
    DEFAULT_CONFIG_NAME = "default"

    def _info(self):
        # Define the features of your dataset
        features = datasets.Features({
            "path": datasets.Value("string"),
            "url": datasets.Value("string"),
            "type": datasets.Value("string"),
            "duration": datasets.Value("float32"),
            "language": datasets.Value("string"),
            "transcript": datasets.Value("string"),
            "tag": datasets.Value("string"),
            "split": datasets.Value("string"),
            "license": datasets.Value("string")
        })
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            license=_LICENSE,
        )

    def _split_generators(self, dl_manager):
        """Split the dataset into train, validation, and test."""
        # Your dataset might have specific splits like "train", "dev", "test"
        splits = ["train", "validation", "test"]

        csv_path = dl_manager.download_and_extract("dataset.csv")

        # Define the splits and pass the language and tag filters to _generate_examples
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"filepath": csv_path, "split": "train", "language": _LANGUAGES, "tag": _TAGS}
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={"filepath": csv_path, "split": "validation", "language": _LANGUAGES, "tag": _TAGS}
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={"filepath": csv_path, "split": "test", "language": _LANGUAGES, "tag": _TAGS}
            ),
        ]


    def _generate_examples(self, csv_file, split, language, tag):
        """Generate examples from the CSV data and audio URLs."""
        # Use dl_manager to get the CSV file from the Hugging Face repo
        # csv_file = dl_manager.download_and_extract("metadata.csv")
        # Read the CSV metadata

        data = pd.read_csv(csv_file)
        # Filter by primary split
        data_split = data[data["split"] == split]

        if language:
            data_split = data_split[data_split["language"] == language]
        if tag:
            data_split = data_split[data_split["tag"] == tag]
        
        for i, row in data_split.iterrows():

            # Construct the full audio path or URL
            audio_url = row["url"]

            # Prepare the data entry for this example
            yield i, {
                "path": row["path"],
                "audio": audio_url,
                "duration": float(row["duration"]),
                "language": row["language"],
                "transcript": row["transcript"],
                "tag": row["tag"],
                "split": row["split"],
                "license": row["license"]
            }



        # with open(csv_file, mode='r', encoding='utf-8') as f:
        #     reader = csv.DictReader(f)
        #     for row in reader:
        #         # Filter by tag or language here if necessary
        #         if row["language"] not in _LANGUAGES:
        #             continue  # Skip entries with unsupported languages
        #         if row["tag"] not in _TAGS:
        #             continue  # Skip entries with unsupported tags




        #         # Construct the full audio path or URL
        #         audio_url = row["url"]

        #         # Prepare the data entry for this example
        #         yield row["path"], {
        #             "audio": audio_url,
        #             "duration": float(row["duration"]),
        #             "language": row["language"],
        #             "transcript": row["transcript"],
        #             "tag": row["tag"],
        #             "split": row["split"],
        #             "license": row["license"]
        #         }



    def _download_audio(self, audio_url):
        """Download audio from a URL if needed (you could also implement streaming)."""
        # This is an example function for downloading audio if it's needed
        # You can integrate this within your data processing pipeline if required
        pass