asr_dataset.py

import os
import tarfile
import pandas as pd
import datasets
from datasets import Audio, Value, Features
import logging
from typing import Dict, Generator, Tuple

logger = logging.getLogger(__name__)


_DESCRIPTION = """
This dataset consists of various Youtube videos in Persian language.

Note: This dataset contains raw, unvalidated transcriptions. Users are advised to:
1. Perform their own quality assessment
2. Create their own train/validation/test splits based on their specific needs
3. Validate a subset of the data if needed for their use case
"""

_CITATION = """
Use this repo info/link for citation.
"""


class ASRDataset(datasets.GeneratorBasedBuilder):
    """ASR dataset with audio files stored in tar archives."""

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features({
                "file_name": datasets.Value("string"),
                "audio": datasets.Audio(sampling_rate=16000),
                "sentence": datasets.Value("string"),
                "tar_file": datasets.Value("string"),
            }),
            supervised_keys=None,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators with added error handling."""
        try:
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,
                    gen_kwargs={
                        "split": "train",
                        "data_dir": self.config.data_dir,
                    },
                ),
            ]
        except Exception as e:
            logger.error(f"Error in _split_generators: {e}")
            logger.error(traceback.format_exc())
            raise

    def _prepare_metadata(self, data_dir: str) -> pd.DataFrame:
        """Prepare metadata with comprehensive error handling."""
        try:
            logger.info("Preparing metadata with tar file information...")
            
            # Read metadata
            metadata_path = os.path.join(data_dir, "metadata.csv")
            
            # Add error handling for metadata reading
            try:
                df = pd.read_csv(metadata_path, sep='\t', names=['file_name', 'sentence'], encoding="utf-8")
            except Exception as read_error:
                logger.error(f"Failed to read metadata file: {read_error}")
                raise
            
            # Validate dataframe
            if df.empty:
                raise ValueError("Metadata file is empty")
            
            # Add tar_file column
            clips_dir = os.path.join(data_dir, "clips")
            tar_files = [f for f in os.listdir(clips_dir) if f.endswith('.tar')]
            
            # Initialize tar_file column
            df['tar_file'] = None
            
            # Find which tar file contains each audio file
            for tar_file in tar_files:
                tar_path = os.path.join(clips_dir, tar_file)
                try:
                    with tarfile.open(tar_path, 'r') as tar:
                        file_list = tar.getnames()
                        mask = df['file_name'].isin(file_list)
                        df.loc[mask, 'tar_file'] = tar_file
                except Exception as tar_error:
                    logger.warning(f"Error processing tar file {tar_file}: {tar_error}")
            
            # Remove entries where tar_file is None (file not found in any tar)
            df = df.dropna(subset=['tar_file'])
            
            # Additional logging
            logger.info(f"Total entries after tar file mapping: {len(df)}")
            
            return df
        
        except Exception as e:
            logger.error(f"Unexpected error in _prepare_metadata: {e}")
            logger.error(traceback.format_exc())
            raise

    def _generate_examples(self, split, data_dir):
        """Yields examples with comprehensive error handling."""
        try:
            # Prepare metadata with tar file information
            df = self._prepare_metadata(data_dir)
            clips_dir = os.path.join(data_dir, "clips")
            
            idx = 0
            # Group by tar_file for efficient processing
            for tar_file, group in df.groupby('tar_file'):
                tar_path = os.path.join(clips_dir, tar_file)
                logger.info(f"Processing tar file: {tar_file}")
                
                try:
                    with tarfile.open(tar_path, 'r') as tar:
                        for _, row in group.iterrows():
                            try:
                                # More robust file extraction
                                member = tar.getmember(row['file_name'])
                                f = tar.extractfile(member)
                                if f is None:
                                    logger.warning(f"Could not extract file: {row['file_name']}")
                                    continue
                                    
                                audio_bytes = f.read()
                                f.close()
                                
                                yield idx, {
                                    "file_name": row['file_name'],
                                    "audio": {"path": f"{tar_path}::{row['file_name']}", "bytes": audio_bytes},
                                    "sentence": row['sentence'],
                                    "tar_file": tar_file,
                                }
                                idx += 1
                                
                            except Exception as file_error:
                                logger.warning(f"Error processing file {row['file_name']}: {file_error}")
                                continue
                                
                except Exception as tar_error:
                    logger.error(f"Error processing tar file {tar_file}: {tar_error}")
                    continue
            
            logger.info(f"Total examples generated: {idx}")
        
        except Exception as e:
            logger.error(f"Unexpected error in _generate_examples: {e}")
            logger.error(traceback.format_exc())
            raise