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)

Tip

If you don’t have uv installed, refer to the Installation Guide for setup instructions, or install it quickly with:

curl -LsSf https://astral.sh/uv/0.8.22/install.sh | sh
source $HOME/.local/bin/env

Installation Options#

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

The simplest way to install NeMo Curator with audio support:

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

Note

The audio extras include NeMo Toolkit with ASR models. Additional audio processing libraries (soundfile, editdistance) are installed automatically as NeMo Toolkit dependencies.

Install the latest version directly from GitHub:

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

Note

Use audio_cpu for CPU-only audio processing, audio_cuda12 for GPU acceleration, or all for all modalities.

NeMo Curator is available as a standalone container:

# Pull the container
docker pull nvcr.io/nvidia/nemo-curator:25.09

# Run the container
docker run --gpus all -it --rm nvcr.io/nvidia/nemo-curator:25.09

See also

For details on container environments and configurations, see Container Environments.

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 -m nemo_curator.examples.audio.fleurs.run \
    --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: