fleurs

Curating the FLEURS Dataset with NeMo Curator

The FLEURS dataset contains spoken utterances across 100+ languages. This tutorial shows how to:

• Download a FLEURS split for a given language
• Run ASR inference using a NeMo model
• Compute WER and audio duration
• Filter by WER threshold
• Save results as JSONL

Note: Run these examples on GPUs for best performance.

Quick start (Python script)

The Python pipeline downloads the data, runs ASR, computes WER/duration, filters, and writes JSONL results to result/ under your raw_data_dir.

python tutorials/audio/fleurs/pipeline.py \
  --raw_data_dir ./example_audio/fleurs \
  --model_name nvidia/stt_hy_fastconformer_hybrid_large_pc \
  --lang hy_am \
  --split dev \
  --wer_threshold 75 \
  --gpus 1 \
  --backend xenna \
  --clean \
  --verbose

Key arguments:

• --raw_data_dir: Workspace directory for downloaded audio and outputs
• --model_name: NeMo ASR model (change per language)
• --lang: FLEURS language code (e.g., hy_am, en_us, etc.)
• --split: FLEURS split (train, dev, or test)
• --wer_threshold: Keep samples with WER less-or-equal to this value
• --backend: xenna (default) or ray_data

Choosing a Backend

Both the Python script (pipeline.py) and the YAML runner (run.py) support two execution backends:

Backend	Description	When to use
xenna	Default executor. Uses Cosmos-Xenna streaming engine with automatic worker allocation.	Most workloads, CI/nightly benchmarks.
ray_data	Executor built on Ray Data map_batches.	Development, machines where Xenna cannot detect GPUs, or when Ray Data integration is preferred.

Python script — pass --backend:

python tutorials/audio/fleurs/pipeline.py \
  --raw_data_dir ./example_audio/fleurs \
  --model_name nvidia/parakeet-tdt-0.6b-v2 \
  --lang en_us --split dev --wer_threshold 75 --gpus 1 \
  --backend ray_data

YAML runner — override backend=:

python tutorials/audio/fleurs/run.py \
  --config-path . --config-name pipeline.yaml \
  raw_data_dir=./example_audio/fleurs \
  backend=ray_data

Alternative: YAML config + Hydra

You can run the same workflow by instantiating stages from a YAML config.

Option 1: Edit pipeline.yaml to set raw_data_dir, then run:

python tutorials/audio/fleurs/run.py \
  --config-path . --config-name pipeline.yaml \
  raw_data_dir=./example_audio/fleurs

Option 2: Override values from the command line without editing the file:

python tutorials/audio/fleurs/run.py \
  --config-path . --config-name pipeline.yaml \
  raw_data_dir=./example_audio/fleurs \
  data_split=dev \
  processors.0.lang=en_us \
  processors.1.model_name=nvidia/stt_en_conformer_ctc_large \
  processors.4.target_value=50.0

Note: --config-path . tells Hydra to look for configs in the same directory as run.py.

Notes on overrides (match indices in processors list inside pipeline.yaml):

• processors.0.lang: language for the FLEURS downloader stage
• processors.1.model_name: NeMo ASR model used for inference
• processors.4.target_value: WER threshold used for filtering
• data_split: top-level variable referenced by the first stage as split
• backend: xenna (default) or ray_data

Output

Results are written as JSONL under ${raw_data_dir}/result. Each line contains fields like:

{"audio_filepath": "relative/path/to/audio.wav", "text": "reference transcription", "duration": 4.21}

Depending on configuration, you may also compute and filter by WER using the predicted text. The example configs keep samples with WER less than or equal to the threshold.

GPU/CPU, cleaning, and performance notes

• ASR inference is GPU-accelerated. The YAML config requests one GPU via processors.1.resources.gpus: 1.0. For CPU fallback with the Python script, pass --gpus 0.
• Use --clean to remove an existing result/ directory before writing outputs.
• Use --verbose for DEBUG-level logs, helpful for intermittent issues.
• Reduce or increase batch sizes by editing pipeline.py or pipeline.yaml (e.g., CreateInitialManifestFleursStage().with_(batch_size=4)).
• Lower-memory GPUs may require smaller batch sizes; high-memory GPUs can use larger ones for higher throughput.

Customizing languages and models

• FLEURS language codes follow the dataset convention (e.g., en_us, fr_fr, hy_arm). See the dataset card for a complete list.
• Use a corresponding NeMo model for your target language. For English, for example: nvidia/stt_en_conformer_ctc_large. For Armenian (shown above): nvidia/stt_hy_fastconformer_hybrid_large_pc.

What the stages do

Both the Python and YAML flows compose the same stages:

• Download/create initial manifest for the requested FLEURS split
• Run ASR inference with a specified NeMo model
• Compute pairwise WER between reference and predicted text
• Compute audio durations
• Filter samples by WER threshold
• Convert to document format and write JSONL

After running, inspect ${raw_data_dir}/result to explore your curated manifest(s).