ganga4364/stt_benchmark

📊 Benchmark STT Datasets

This repository contains benchmark Speech-to-Text (STT) datasets designed for fair and consistent evaluation of STT models. The datasets are hosted on Hugging Face, and each one is carefully curated to ensure balanced sampling across departments and attributes. Below, you’ll find links to the datasets, a summary of their benchmarks, and details about the methodology used for their preparation.

---

📂 Datasets Overview

Dataset	Description	Link
AB	Audiobook dataset with attributes like age, gender, and music.	AB Dataset
CS	Children speech recordings with features like speaker ID and microphone type.	CS Dataset
HS	Historical speeches dataset with details like publishing year and place of origin.	HS Dataset
MV	Movie audio data with attributes like subtitles.	MV Dataset
NS	Natural speech recordings with attributes like location and speaker names.	NS Dataset
NW	News dataset with features like media organization and audio quality.	NW Dataset
PC	Podcast recordings with attributes like publishing year and gender.	PC Dataset
TT	Tibetan Teachings and talks dataset with topics and episodes.	TT Dataset

---

📝 Benchmark Summary

Each dataset contains ~10,000 samples, distributed proportionally among its attributes and categories. For categories or speakers with limited data, up to 30% of their data was used to maintain a balance between benchmark and training datasets.

AB (Audiobook)

• Age: 1680 rows
• Gender: 2000 rows
• Music: 2000 rows
• Name: 1607 rows

CS (Children speech)

• Recording Type: 1707 rows
• Class: 319 rows
• Gender: 1526 rows
• Speakers ID: 1807 rows
• Microphone Type: 1675 rows
• School: 1316 rows

NS (Natural speech)

• Mic/Phone: 1526 rows
• Location: 1194 rows
• Speaker Name: 1974 rows
• Gender: 1901 rows
• Age: 1925 rows

PC (Podcast)

• Name of Podcast: 1915 rows
• Publishing Year: 1901 rows
• Channel: 1460 rows
• Audio Quality: 2000 rows
• Country: 1332 rows
• Gender: 1611 rows
• Age Group: 1229 rows

HS (Historical Speeches)

• Publishing Year: 1968 rows
• Name: 1970 rows
• Gender: 2000 rows
• Age Group: 1397 rows
• Place of Origin: 1363 rows
• Exiled Year: 1236 rows

NW (News)

• Name of News: 1901 rows
• Publishing Year: 1755 rows
• Media Organization: 1374 rows
• Audio Quality: 1710 rows
• Single Speaker: 2000 rows
• Country: 1123 rows

MV (Movies)

• Name of the Movie: 1776 rows
• Subtitle: 1491 rows

TT (Tibetan Teachings & Talks)

• Teacher: 1995 rows
• Topic: 1888 rows
• Books: 348 rows
• Episode: 1831 rows

---

🔍 Methodology

The benchmarks were created with the following principles:

1. Proportional Sampling: Approximately 10,000 samples per department, distributed equally across unique values in each column.
2. Fair Representation: Data from each unique value (e.g., speaker, location) was proportionally sampled. For categories with limited data, up to 30% of the available data was used.
3. Consistency: Balanced sampling was maintained to avoid bias in benchmark datasets while preserving the majority of the data for training.
4. Exclusions: Categories or speakers with extremely low data availability were excluded.

---

💾 How to Download Datasets

You can download all datasets programmatically using the script below:

from datasets import load_dataset
import json

# Load dataset metadata
datasets = {
    "AB": "ganga4364/benchmark-stt-AB",
    "CS": "ganga4364/benchmark-stt-CS",
    "HS": "ganga4364/benchmark-stt-HS",
    "MV": "ganga4364/benchmark-stt-MV",
    "NS": "ganga4364/benchmark-stt-NS",
    "NW": "ganga4364/benchmark-stt-NW",
    "PC": "ganga4364/benchmark-stt-PC",
    "TT": "ganga4364/benchmark-stt-TT"
}

# Download and save datasets
for name, url in datasets.items():
    print(f"Downloading {name} dataset...")
    dataset = load_dataset(url)
    dataset.save_to_disk(f"./downloaded_datasets/{name}")

---