Get Started with Audio Curation

This guide helps you set up and get started with NeMo Curator’s audio curation capabilities. Follow these steps to prepare your environment and run your first audio curation pipeline using the FLEURS dataset.

Prerequisites

To use NeMo Curator’s audio curation modules, ensure you meet the following requirements:

• Python 3.10, 3.11, or 3.12
  • packaging >= 22.0
• uv (for package management and installation)
• Ubuntu 22.04/20.04
• NVIDIA GPU (recommended for ASR inference)
  • Volta™ or higher (compute capability 7.0+)
  • CUDA 12 (or above)
• Audio processing libraries (automatically installed with audio extras)

Installation Options

You can install NeMo Curator with audio support in four ways:

echo "transformers==4.55.2" > override.txt
uv pip install "nemo-curator[audio_cuda12]" --override override.txt

git clone https://github.com/NVIDIA-NeMo/Curator.git
cd Curator
uv sync --extra audio_cuda12 --all-groups
source .venv/bin/activate

# Pull the container
docker pull nvcr.io/nvidia/nemo-curator:`container_version`

# Run the container
docker run --gpus all -it --rm nvcr.io/nvidia/nemo-curator:`container_version`

Download Sample Configuration

NeMo Curator provides a sample FLEURS configuration for audio curation. You can download and customize it:

mkdir -p ~/nemo_curator/configs
wget -O ~/nemo_curator/configs/fleurs_pipeline.yaml https://raw.githubusercontent.com/NVIDIA/NeMo-Curator/main/tutorials/audio/fleurs/pipeline.yaml

This configuration file contains a complete audio curation pipeline for the FLEURS dataset, including ASR inference, quality assessment, and filtering.

Set Up Data Directory

Create a directory to store your audio datasets:

mkdir -p ~/nemo_curator/audio_data

Basic Audio Curation Example

Here’s a simple example to get started with audio curation using the FLEURS dataset:

from nemo_curator.pipeline import Pipeline
from nemo_curator.stages.audio.datasets.fleurs.create_initial_manifest import CreateInitialManifestFleursStage
from nemo_curator.stages.audio.inference.asr_nemo import InferenceAsrNemoStage
from nemo_curator.stages.audio.metrics.get_wer import GetPairwiseWerStage
from nemo_curator.stages.audio.common import GetAudioDurationStage, PreserveByValueStage
from nemo_curator.stages.resources import Resources

# Create audio curation pipeline
pipeline = Pipeline(name="audio_curation", description="FLEURS audio curation with ASR and WER filtering")

# 1. Load FLEURS dataset (Armenian development set)
pipeline.add_stage(
    CreateInitialManifestFleursStage(
        lang="hy_am",
        split="dev",
        raw_data_dir="~/nemo_curator/audio_data"
    ).with_(batch_size=4)
)

# 2. Perform ASR inference using NeMo model
pipeline.add_stage(
    InferenceAsrNemoStage(
        model_name="nvidia/stt_hy_fastconformer_hybrid_large_pc"
    ).with_(resources=Resources(gpus=1.0))
)

# 3. Calculate Word Error Rate (WER)
pipeline.add_stage(
    GetPairwiseWerStage(
        text_key="text",
        pred_text_key="pred_text",
        wer_key="wer"
    )
)

# 4. Calculate audio duration
pipeline.add_stage(
    GetAudioDurationStage(
        audio_filepath_key="audio_filepath",
        duration_key="duration"
    )
)

# 5. Filter by WER threshold (keep samples with WER <= 75%)
pipeline.add_stage(
    PreserveByValueStage(
        input_value_key="wer",
        target_value=75.0,
        operator="le"  # less than or equal
    )
)

# Execute the pipeline
pipeline.run()

Alternative: Configuration-Based Approach

You can also run the pipeline using the downloaded configuration:

cd ~/nemo_curator
python tutorials/audio/fleurs/run.py \
    --config-path ~/nemo_curator/configs \
    --config-name fleurs_pipeline.yaml \
    raw_data_dir=~/nemo_curator/audio_data

Expected Output

After running the pipeline, you’ll have:

~/nemo_curator/audio_data/
├── hy_am/                    # Armenian language data
│   ├── dev.tsv              # Transcription metadata
│   ├── dev.tar.gz          # Audio archive
│   ├── dev/                # Extracted audio files
│   └── result/             # Filtered results
│       └── *.jsonl        # High-quality audio-text pairs

Each output entry contains:

{
    "audio_filepath": "/absolute/path/to/audio.wav",
    "text": "ground truth transcription",
    "pred_text": "asr model prediction",
    "wer": 12.5,
    "duration": 4.2
}

Next Steps

Explore the Audio Curation documentation for more advanced processing techniques and customization options.

Key areas to explore next:

• Custom Audio Manifests - Load your own audio datasets
• Quality Assessment - Advanced filtering and quality metrics
• Text Integration - Combine with text processing workflows