espnet2.speechlm.dataloader.multimodal_loader.audio_loader.OmniIOAudioReader

source

class espnet2.speechlm.dataloader.multimodal_loader.audio_loader.OmniIOAudioReader(parquet_path: str, valid_ids: list | None = None)

Bases: object

Dict-like lazy audio reader using omniio parquets.

Reads audio data from omniio parquet files. Audio is accessed via byte offsets and time boundaries stored in the parquet metadata.

• Returns: Tuple of (audio_array, sample_rate) where audio_array has shape [num_samples, num_channels].
• Parameters:
  • parquet_path – Path to the parquet file containing audio metadata.
  • valid_ids – List of valid IDs to keep (optional, keeps all if None).

items() → Iterator[Tuple[str, Tuple[ndarray, int]]]

Return iterator over (id, (audio_array, sample_rate)) pairs.

keys() → Iterator[str]

Return iterator over IDs.

values() → Iterator[Tuple[ndarray, int]]

Return iterator over (audio_array, sample_rate) tuples.