Save & Export Audio Data

Export processed audio data and transcriptions in formats optimized for ASR model training, audio-and-text applications, and downstream analysis workflows.

Output Formats

NeMo Curator’s audio curation pipeline supports several output formats tailored for different use cases:

JSONL Manifests

The primary output format for audio curation is JSONL (JSON Lines):

1
{"audio_filepath": "/data/audio/sample_001.wav", "text": "hello world", "pred_text": "hello world", "wer": 0.0, "duration": 2.1}
2
{"audio_filepath": "/data/audio/sample_002.wav", "text": "good morning", "pred_text": "good morning", "wer": 0.0, "duration": 1.8}

Metadata Fields

Standard fields included in audio manifests:

Export Configuration

1
from nemo_curator.stages.text.io.writer import JsonlWriter
2
from nemo_curator.stages.audio.io.convert import AudioToDocumentStage
3
4
# Convert AudioBatch to DocumentBatch for text writer
5
pipeline.add_stage(AudioToDocumentStage())
6
7
# Configure JSONL export
8
pipeline.add_stage(
9
    JsonlWriter(
10
        path="/output/audio_manifests",
11
        write_kwargs={"force_ascii": False}  # Support Unicode characters
12
    )
13
)

Directory Structure

Standard Output Layout

When source_files metadata exists, the writer generates deterministic hashed file names. Otherwise, it generates UUID-based names.

/output/audio_manifests/
├── <hash>.jsonl   # Deterministic hash if metadata.source_files present, else UUID
├── <hash>.jsonl
└── ...

Quality Control

Validation Checks

Before export, check your processed data:

1
from nemo_curator.stages.audio.common import PreserveByValueStage
2
3
# Filter by quality thresholds
4
quality_filters = [
5
    # Keep samples with WER <= 50%
6
    PreserveByValueStage(
7
        input_value_key="wer",
8
        target_value=50.0,
9
        operator="le"
10
    ),
11
    # Keep samples with duration 1-30 seconds
12
    PreserveByValueStage(
13
        input_value_key="duration", 
14
        target_value=1.0,
15
        operator="ge"
16
    ),
17
    PreserveByValueStage(
18
        input_value_key="duration",
19
        target_value=30.0, 
20
        operator="le"
21
    )
22
]
23
24
for filter_stage in quality_filters:
25
    pipeline.add_stage(filter_stage)