Kaushalb11/multi-lang-s2t-dataset

Multi-Language Audio Dataset

A high-quality, cleaned audio dataset derived from five AI4Bharat sources, containing speech-to-text data in Hindi, Gujarati, and Telugu with English translations.

Dataset Overview

Language	Duration (Hours)
Hindi	55.45
Gujarati	42.12
Telugu	37.83

Total Duration: ~135 hours of cleaned, high-quality audio

Source Datasets

Data collected and cleaned from:

• ai4bharat/IndicVoices-ST
• ai4bharat/NPTEL
• ai4bharat/WordProject
• ai4bharat/Spoken-Tutorial
• ai4bharat/Mann-ki-Baat

Quality Metrics

All audio samples passed rigorous quality checks with a 0.75+ threshold score:

• SNR Estimation: Signal-to-noise ratio analysis
• RMS dB Levels: Volume consistency checks
• Silence Detection: Minimal silence ratios
• Clipping Detection: No audio distortion
• DC Offset Correction: Centered waveforms
• Spectral Analysis: Frequency distribution validation

Dataset Structure

multi_language_perfect_audio/
├── gujarati_perfect_audio/
│   ├── gujarati_part_001.tar.zst
│   ├── gujarati_part_002.tar.zst
├── hindi_perfect_audio/
│   ├── hindi_part_001.tar.zst
│   ├── hindi_part_002.tar.zst
│   ├── hindi_part_003.tar.zst
├── telugu_perfect_audio/
│   ├── telugu_part_001.tar.zst
│   ├── telugu_part_002.tar.zst
├── sample_audio/
│   ├── hindi/
│   │   ├── *.wav
│   ├── gujarati/
│   ├── telugu/
├── gujarati_summary.json
├── hindi_summary.json
├── telugu_summary.json
├── gujarati_perfect_audio.csv
├── hindi_perfect_audio.csv
└── telugu_perfect_audio.csv

File Descriptions

Audio Files

• Compressed archives (.tar.zst) containing perfect-quality audio samples
• Format: WAV files, 16kHz sample rate
• Naming: {dataset_name}_{language}_{sample_id}.wav

Metadata Files

{lang}_perfect_audio.csv

Contains metadata for all perfect-quality audio samples including transcriptions in native language, English translations, audio paths, duration, and quality metrics (SNR, RMS, quality score).

{lang}_summary.json

Comprehensive statistics including total samples examined, valid/perfect sample counts, duration metrics, noise analysis, and overall quality percentages for each language.

Sample Data

We provide 10 sample audio files per language (2 samples from each of the 5 source datasets) in the sample_audio/ directory for preview and testing purposes.