NeMo ASR Models

Use NeMo Framework’s automatic speech recognition models for transcription in your audio curation pipelines. This guide covers basic usage and configuration.

Model Selection

NeMo Framework provides pre-trained ASR models through the Hugging Face model hub. For the complete list of available models and their specifications, refer to the NeMo Framework ASR documentation.

Example Model Usage

# Example using a test-verified model
example_model = "nvidia/parakeet-tdt-0.6b-v2"
# For production use, select appropriate models from:
# https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/all_chkpt.html

Basic Usage

Simple ASR Inference

from nemo_curator.stages.audio.inference.asr.stage import ASRStage
from nemo_curator.stages.resources import Resources
# Create ASR inference stage with a model from NeMo Framework
asr_stage = ASRStage(
    adapter_target="nemo_curator.models.asr.nemo_asr.NeMoASRAdapter",
    model_id="nvidia/parakeet-tdt-0.6b-v2",
    max_audio_sec_per_actor=240.0,
    max_inference_duration_s=120.0,
    local_bucketing=True,
    audio_filepath_key="audio_filepath",
    pred_text_key="pred_text",
    batch_size=16,
)
# Configure for GPU processing
asr_stage = asr_stage.with_(
    resources=Resources(gpus=1.0),
)

Custom Configuration

# Example with custom field names
custom_asr = ASRStage(
    adapter_target="nemo_curator.models.asr.nemo_asr.NeMoASRAdapter",
    model_id="nvidia/parakeet-tdt-0.6b-v2",
    max_audio_sec_per_actor=240.0,
    max_inference_duration_s=120.0,
    local_bucketing=True,
    audio_filepath_key="custom_audio_path",
    pred_text_key="transcription",
    batch_size=32,
).with_(
    resources=Resources(cpus=4.0, gpus=1.0),
)

Model Caching

The executor downloads and caches model weights once per node, then loads one adapter in each worker:

asr_stage = ASRStage(
    adapter_target="nemo_curator.models.asr.nemo_asr.NeMoASRAdapter",
    model_id="nvidia/parakeet-tdt-0.6b-v2",
    max_audio_sec_per_actor=240.0,
    max_inference_duration_s=120.0,
    local_bucketing=True,
    audio_filepath_key="audio_filepath",
)
# Add the stage to a Pipeline and run it through an executor. Do not call
# setup() or process_batch() directly; the executor owns worker lifecycle.

Resource Configuration

Configure GPU and CPU resources based on your hardware:

from nemo_curator.stages.resources import Resources
# Single GPU configuration
asr_stage = ASRStage(
    adapter_target="nemo_curator.models.asr.nemo_asr.NeMoASRAdapter",
    model_id="nvidia/parakeet-tdt-0.6b-v2",
    max_audio_sec_per_actor=240.0,
    max_inference_duration_s=120.0,
    local_bucketing=True,
    audio_filepath_key="audio_filepath",
    batch_size=16,
).with_(
    resources=Resources(
        cpus=4.0,
        gpus=1.0,
    ),
)

NeMoASRAdapter uses zero or one GPU per worker. Scale throughput by allowing the executor to run more one-GPU workers instead of assigning multiple GPUs to one stage instance.

Resource requirements vary by model. Test with your specific model to determine optimal settings.