hoangbang/speak-the-digit

Speak the Digit: Spoken Digit Recognition

Dataset Summary

A public, viewer-ready educational challenge dataset. Host-only scoring data and hidden targets are excluded.

Splits

Split	Examples	Description
train	2,400	Labeled training data
test	600	Public inputs with withheld target labels or annotations

Data Fields

Field	Type
audio	Audio
id	string
label	string (test sentinel: unlabeled)
duration_seconds	float64
sampling_rate	int64

Labels

• 0
• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9

Viewer and Loading

The standard repository layout enables the Hugging Face Dataset Viewer after processing. Load with the datasets library using hoangbang/speak-the-digit.

from datasets import load_dataset

dataset = load_dataset("hoangbang/speak-the-digit")

Preparation

• Tabular data is stored as Parquet.
• Image classification uses class directories and ImageFolder metadata.
• Object detection uses pixel-space COCO xywh metadata converted from YOLO annotations.
• Audio uses WAV folders and split metadata compatible with AudioFolder.
• ground_truth.csv, host_only, filename maps, source archives, and local OS files are excluded.
• The literal __unlabeled__ in the test label field is a no-label sentinel, not an annotation.

Source, License, and Attribution

• Source: https://www.kaggle.com/datasets/joserzapata/free-spoken-digit-dataset-fsdd
• License note: See original source terms and attribution before reuse.
• Check original source terms before further redistribution or commercial use.

Limitations and Responsible Use

• Limited speakers and recording conditions.
• Test labels are intentionally withheld.