Datasetpublic

Peacockery/farsi-asr-wer35-fastconformer

Farsi ASR WER35 FastConformer This dataset contains Farsi/Farsi ASR utterances curated with NVIDIA NeMo Curator using nvidia/stt_fa_fastconformer_hybrid_large. The uploaded training data is stored as WebDataset TAR shards because Hugging Face recommends WebDataset archives for large-scale audio datasets. The local artifact also includes a NeMo ASR manifest at manifests/train_manifest.jsonl. Curation Language: Farsi/Farsi (fa) Audio: FLAC, 16 kHz mono Source run:… See the full description on the dataset page: https://huggingface.co/datasets/Peacockery/farsi-asr-wer35-fastconformer.

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes5downloads

Dataset Card

Farsi ASR WER35 FastConformer

This dataset contains Farsi/Farsi ASR utterances curated with NVIDIA NeMo Curator using nvidia/stt_fa_fastconformer_hybrid_large.

The uploaded training data is stored as WebDataset TAR shards because Hugging Face recommends WebDataset archives for large-scale audio datasets. The local artifact also includes a NeMo ASR manifest at manifests/train_manifest.jsonl.

Curation

—Language: Farsi/Farsi (fa)
—Audio: FLAC, 16 kHz mono
—Source run: full-p018-dataset-fa
—Candidate rows before Curator filtering: 508,518
—Kept rows: 416,056
—Kept hours: 500.51
—WER model: nvidia/stt_fa_fastconformer_hybrid_large
—WER threshold: <= 35
—Duration range: 1.0 <= duration <= 20.0 seconds
—Mean WER after filtering: 5.1240

Source Counts

json

{
  "common_voice_17_0": 389393,
  "thomcles_farsi_farsi_speech": 24407,
  "fleurs": 2256
}

Files

—webdataset/train-*.tar: audio and per-sample JSON metadata, WebDataset format
—manifests/train_manifest.jsonl: NeMo ASR manifest for local fine-tuning
—metadata/metadata.jsonl: compact row metadata for auditing
—metadata/metadata.parquet: Parquet copy of row metadata for analysis
—summary.json: counts, thresholds, provenance, shard checksums

Verification

This artifact was built from NeMo Curator output and checked for:

—exactly 416,056 unique sample_id values
—every referenced FLAC exists locally while building
—every row has wer <= 35
—every row has 1.0 <= duration <= 20.0
—source counts and total hours written to summary.json

References

—Hugging Face audio dataset layout and WebDataset guidance: https://huggingface.co/docs/hub/en/datasets-audio
—Hugging Face audio dataset creation guide: https://huggingface.co/docs/datasets/audio_dataset
—NVIDIA NeMo ASR manifest format: https://docs.nvidia.com/nemo-framework/user-guide/latest/nemotoolkit/asr/datasets.html
—NVIDIA model used for WER scoring: https://huggingface.co/nvidia/sttfafastconformerhybridlarge