jenil17/Swayam_speech_dataset

SWAYAM English Lecture Speech Dataset

📌 Overview

The SWAYAM English Lecture Speech Dataset is a curated speech dataset extracted from publicly available SWAYAM course lectures.

The dataset consists of English speech segments derived from academic lecture recordings across five university-level courses.

It is intended for:

• Automatic Speech Recognition (ASR)
• Speech-to-Text (S2T)
• Lecture transcription research
• Educational speech modeling
• Domain-specific speech adaptation

---

🎓 Source Courses

The dataset is constructed from the following SWAYAM courses:

1. Access to Justice
2. Biomolecules: Structure & Functions in Health and Disease
3. Corporate Tax Planning
4. MOOC on FMFS – Dr. Tejpal Dhewa (UGC)
5. Solid & Hazardous Waste Management

All lectures are in English.

---

📊 Dataset Statistics

📁 Course-wise Statistics

Course Name	Total Files	Valid Files	Corrupted Files	Total Duration (Hours)	Avg Segment Length (Sec)
Access to Justice	5,471	5,471	0	18.91	12.44
Biomolecules Structure & Functions	7,224	6,739	485	20.94	11.19
Corporate Tax Planning	2,305	2,305	0	8.18	12.78
MOOC on FMFS (UGC)	4,970	4,749	221	15.48	11.73
Solid & Hazardous Waste Management	4,631	4,631	0	16.23	12.62
Total	24,601	23,895	706	79.75+ Hours	~12.01

---

📂 Dataset Structure

The dataset is organized by course. Each directory contains segmented English lecture audio files in .wav format.

final_dataset/

• access_to_justice_multilingual_speech/
• biomolecules_multilingual_speech_dataset/
• Corporate-Tax-Planning-Dataset/
• MOOC-on_FMFS_Dr-Tejpal-Dhewa_UGC-Dataset/
• Soild_Hardagious_waste_Mgmt-Dataset/

Each subfolder represents one SWAYAM course and contains segmented lecture audio clips.

---

🔍 Data Quality Summary

Total Audio Files: 24,601
Valid Files: 23,895
Empty Files: 0
Corrupted Files: 706
Total Duration: ~79.75 Hours
Average Segment Duration (Overall): ~12.01 seconds

All files were programmatically validated using automated audio metadata inspection.

---

🎧 Audio Format

• Format: WAV
• Language: English
• Domain: Academic Lecture Speech
• Average Segment Duration: ~10–13 seconds

---

🛠 Intended Use

This dataset can be used for:

• ASR training
• Domain adaptation for lecture transcription
• Academic speech recognition benchmarking
• Speech segmentation research
• Long-form lecture modeling

---

⚠ Limitations

• No official transcripts are included in this release.
• Dataset is domain-specific (academic lectures).
• 706 files were identified as corrupted. Users may filter these files if required.
• Audio is segmented from long-form lecture recordings.

---

📜 License

This dataset is derived from publicly available SWAYAM educational lecture material.

License: CC-BY-4.0

Users must comply with original content licensing policies.

---

🙏 Acknowledgments

LTRC, IIIT Hyderabad for providing GPU facilities and mentorship.