Class LibrispeechLoader

scripts/eval/eval.py:320–371

Dataset loader for LibriSpeech test sets. LibriSpeech is a large corpus of approximately 1000 hours of English speech derived from LibriVox audiobooks. This loader handles both 'test-clean' and 'test-other' evaluation subsets. Dataset Structure: - Audio files in FLAC format

class LibrispeechLoader(BaseDatasetLoader):
    """Dataset loader for LibriSpeech test sets.

    LibriSpeech is a large corpus of approximately 1000 hours of English speech
    derived from LibriVox audiobooks. This loader handles both 'test-clean' and
    'test-other' evaluation subsets.

    Dataset Structure:
        - Audio files in FLAC format organized by speaker/chapter
        - Transcript files (.txt) contain utterance ID and normalized text
        - Standard format: SPEAKER-CHAPTER-UTTERANCE_ID transcript_text

    Returns:
        Tuple of audio file paths and corresponding transcript texts

    Reference:
        Panayotov, V., et al. "Librispeech: an ASR corpus based on public domain audio books."
    """

    def load(self) -> Tuple[list, list]:
        """Load LibriSpeech audio files and transcripts.

        Recursively searches for transcript files and maps them to corresponding
        FLAC audio files using the LibriSpeech naming convention.

        Returns:
            Tuple[list, list]: A tuple containing:
                - List of audio file paths (FLAC format)
                - List of corresponding transcript strings
        """
        transcript_files = []
        audio_text = {}

        for root, _, files in os.walk(self.root_dir):
            transcript_files.extend(
                os.path.join(root, file) for file in files if file.endswith(".txt")
            )

        for file in sorted(transcript_files):
            with open(file, "r") as f:
                for line in f:
                    parts = line.split(" ")
                    audio_codes = parts[0].split("-")
                    audio_file = os.path.join(
                        self.root_dir,
                        audio_codes[0],
                        audio_codes[1],
                        f"{audio_codes[0]}-{audio_codes[1]}-{audio_codes[2]}.flac",
                    )
                    audio_text[audio_file] = " ".join(parts[1:]).strip()

        return list(audio_text.keys()), list(audio_text.values())


class ArtieBiasCorpusLoader(BaseDatasetLoader):