Curate AudioLoad Data

Load Audio Data

View as Markdown

Import audio datasets from various sources into NeMo Curator’s audio processing pipeline. Audio data loading supports manifest files, direct file paths, and automated dataset downloads.

How it Works

Audio data loading in NeMo Curator centers around the AudioBatch data structure, which contains:

  • Audio file paths: References to audio files (.wav, .mp3, .flac, etc.)
  • Transcriptions: Ground truth or reference text for speech content
  • Metadata: Duration, language, speaker information, and quality metrics

The loading process validates audio file existence and formats data for downstream ASR inference and quality assessment stages.

Loading Methods

Choose the appropriate loading method based on your data source and format:

FLEURS Dataset

Automated download and processing of the multilingual FLEURS speech dataset

Custom Manifests

Create and load custom audio manifests with file paths and transcriptions

Local Files

Load audio files directly from local directories and file systems