Class SwitchBoardLoader

scripts/eval/eval.py:719–803

Dataset loader for Switchboard corpus. Switchboard contains conversational telephone speech between strangers discussing predetermined topics. Like CallHome, it uses the HUB5 evaluation format but represents more structured conversational speech. Dataset Structure: - HUB5 S

class SwitchBoardLoader(BaseDatasetLoader):
    """Dataset loader for Switchboard corpus.

    Switchboard contains conversational telephone speech between strangers discussing
    predetermined topics. Like CallHome, it uses the HUB5 evaluation format but
    represents more structured conversational speech.

    Dataset Structure:
        - HUB5 STM format with timing and channel information
        - SPHERE audio files requiring channel extraction
        - Topic-guided conversational speech

    Returns:
        Tuple of AudioSegment objects with timing and corresponding transcript texts

    Reference:
        Godfrey, J. J., et al. "SWITCHBOARD: telephone speech corpus for research and development."
    """

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

        Delegates to the shared HUB5 data loading logic with Switchboard prefix
        to extract Switchboard-specific segments.

        Returns:
            Tuple[list, list]: A tuple containing:
                - List of AudioSegment tuples (wav_file, start_time, end_time)
                - List of corresponding transcript strings
        """
        return self._load_hub5_data("sw")

    def _load_hub5_data(self, prefix: str) -> Tuple[list, list]:
        """Common logic for CallHome and SwitchBoard datasets.

        Parses HUB5 STM format files to extract audio segments with precise timing
        information. Handles channel separation using Sox audio processing.

        Args:
            prefix (str): Dataset prefix ("en" for CallHome, "sw" for Switchboard)

        Returns:
            Tuple[list, list]: A tuple containing:
                - List of AudioSegment tuples (wav_file, start_time, end_time)
                - List of corresponding transcript strings

        Note:
            Requires Sox audio processing tool for SPHERE format conversion
            and channel extraction.
        """
        audio_files, transcript_files = [], []

        stm_path = f"{self.root_dir}/2000_hub5_eng_eval_tr/reference/hub5e00.english.000405.stm"
        with open(stm_path, "r") as f:
            for line in f:
                if line.startswith(";;") or not line.startswith(prefix):
                    continue