Audio Curation Pipeline (Overview)

This guide provides an overview of the end-to-end audio curation workflow in NVIDIA NeMo Curator. It covers data ingestion and validation, optional ASR inference, quality assessment, filtering, and export or conversion. For detailed ASR pipeline information, refer to ASR Pipeline.

High-Level Flow

Core Components

Data Ingestion and Validation:

• AudioTask file existence checks using validate() and validate_item()
• Manifest format validation and metadata consistency
• Recommended JSONL manifest format

Optional ASR Inference:

• InferenceAsrNemoStage for automatic speech recognition
• Configurable batch processing with batch_size and resources parameters
• Support for multiple NeMo ASR models

Quality Assessment:

• Audio duration analysis with GetAudioDurationStage
• Word Error Rate (WER) and Character Error Rate (CER) calculation
• Speech rate metrics including words per second and characters per second

Filtering and Quality Control:

• Threshold-based filtering using PreserveByValueStage
• Configurable quality thresholds for WER, duration, and speech rate

Export and Format Conversion:

• Audio-to-text conversion with AudioToDocumentStage
• Integration with text processing workflows

Common Workflows

ASR-First Workflow (Most Common):

1. Load audio files into AudioTask format
2. Apply ASR inference to generate transcriptions
3. Calculate quality metrics (WER, duration, speech rate)
4. Apply threshold-based filtering
5. Convert to DocumentBatch for text processing integration
6. Export filtered, high-quality audio-text pairs

Quality-First Workflow (No ASR Required):

1. Load audio files with existing transcriptions
2. Extract audio characteristics (duration, format, sample rate)
3. Apply basic quality filters
4. Export validated audio dataset