Datasetpublic

toth235a/voxpopuli-qc-samples-v3

VoxPopuli QC Samples V3 - CER-based Quality Control Quality control samples from VoxPopuli French ASR pseudolabeling, categorized by Character Error Rate (CER) between Whisper (original) and Parakeet (new) transcriptions. View in HuggingFace Dataset Viewer This dataset is viewable directly in the HuggingFace dataset viewer! Click the "Dataset Viewer" tab above to: Listen to audio samples See full Whisper and Parakeet transcriptions (not truncated) Filter by CER… See the full description on the dataset page: https://huggingface.co/datasets/toth235a/voxpopuli-qc-samples-v3.

sourceHugging Facecc-by-4.0updated 9mo agoView on Hugging Face

0likes134downloads

Dataset Card

VoxPopuli QC Samples V3 - CER-based Quality Control

Quality control samples from VoxPopuli French ASR pseudolabeling, categorized by Character Error Rate (CER) between Whisper (original) and Parakeet (new) transcriptions.

View in HuggingFace Dataset Viewer

This dataset is viewable directly in the HuggingFace dataset viewer! Click the "Dataset Viewer" tab above to:

—Listen to audio samples
—See full Whisper and Parakeet transcriptions (not truncated)
—Filter by CER bin

Interactive Notebook (Colab)

![Open In Colab](https://colab.research.google.com/github/toth235a/voxpopuli-qc-samples-v3/blob/main/qcviewercolab.ipynb)

Dataset Summary

—60 samples total (10 per CER bin)
—Columns: audio, id, cer_bin, cer, whisper_text, parakeet_text, duration

CER Bins

Bin	CER Range	Description
`cer_02_05`	2-5%	Good agreement (reference)
`cer_05_10`	5-10%	Slight difference
`cer_10_20`	10-20%	Moderate difference
`cer_20_30`	20-30%	Significant difference
`cer_30_50`	30-50%	Very different
`cer_50_plus`	50%+	Extremely different

CER Distribution (from 2.4M filtered samples)

CER Range	% of Dataset
0-2%	27.8%
2-5%	27.9%
5-10%	22.7%
10-20%	12.8%
20-30%	3.2%
30-50%	2.0%
50%+	2.7%

Mean CER: 20% | Median CER: 4.3%

Usage

python

from datasets import load_dataset

ds = load_dataset("toth235a/voxpopuli-qc-samples-v3", split="train")

for sample in ds:
    print(f"ID: {sample['id']}, CER: {sample['cer']:.1%}")
    print(f"Whisper: {sample['whisper_text']}")
    print(f"Parakeet: {sample['parakeet_text']}")
    # sample['audio'] contains audio array and sampling_rate

Source

—Original transcriptions: Whisper
—New transcriptions: Parakeet TDT 0.6B v3
—Parent dataset: VoxPopuli

License

CC-BY-4.0 (following VoxPopuli license)