surindersinghssj/gurbani-kirtan-dataset-v2

Gurbani Kirtan Dataset V2

A timestamped Gurbani Kirtan dataset with line-level audio segments, canonical Gurmukhi text (matched against SikhiToTheMax database), and English translations extracted from YouTube kirtan videos.

Dataset Description

Each row represents a single Gurbani line (slide) from a kirtan video, with:

• Audio segment (FLAC, 16kHz mono, one per slide/pangati)
• Gurmukhi text — canonical text matched against STTM database
• Gurmukhi OCR — raw OCR output for comparison
• English translation — from Dr. Sant Singh Khalsa (via STTM)
• Slide image — original text overlay from the video
• Kirtan style — studio, gurdwara_live, akj, etc.

Dataset Statistics

Metric	Value
Total segments	9319
Total audio duration	248775.4s (69.1h)
Number of videos	436
STTM matched	9197/9319 (99%)
Avg segment duration	26.7s
Audio format	FLAC, 16kHz, mono
Languages	Punjabi (Gurmukhi), English
Kirtan styles	studio

Splits

Split	Segments	Note
train	8766	95% of videos
validation	227	2.5% of videos
test	326	2.5% of videos

Splits are by video ID (not segment) to prevent data leakage.

Data Fields

Field	Type	Description
audio	Audio	FLAC audio segment (16kHz mono)
gurmukhi_text	string	Canonical Gurbani line (STTM-corrected, cleaned)
gurmukhi_ocr	string	Raw OCR output from slide
gurmukhi_vishram	string	Canonical text with vishram (pause) markers
english_translation	string	English translation (Dr. Sant Singh Khalsa)
match_score	float	STTM fuzzy match confidence (0-100)
start_time	float	Start time in source video (seconds)
end_time	float	End time in source video (seconds)
duration	float	Duration of audio segment (seconds)
slide_index	int	Index of the slide in the video
video_id	string	YouTube video ID
shabad_title	string	Title of the shabad
channel	string	YouTube channel name
kirtan_style	string	Style: studio, gurdwara_live, akj, rain_sabai
segment_type	string	Audio type: vocal, instrumental, or silent
vocal_ratio	float	Fraction of frames with vocal characteristics

How to Use

from datasets import load_dataset

ds = load_dataset("surindersinghssj/gurbani-kirtan-dataset-v2")

# Access splits
train = ds["train"]
val = ds["validation"]
test = ds["test"]

# Filter high-confidence matches only
high_quality = train.filter(lambda x: x["match_score"] >= 80)

Pipeline

1. Download kirtan videos from YouTube using yt-dlp
2. Detect slide transitions using OpenCV frame differencing (auto-threshold: median + 4*std)
3. Extract audio segments between transitions using ffmpeg (FLAC 16kHz mono)
4. OCR slide text using Surya OCR (with Tesseract fallback)
5. Match against STTM database (141K lines) for canonical Gurmukhi + English translations
6. Clean text — remove verse markers (॥), digits, ਰਹਾਉ; filter empty rows
7. Assign train/val/test splits by video ID (95/2.5/2.5)

Source

Videos sourced from YouTube kirtan channels that display each Gurbani line as a slide synchronized to the audio.

License

CC-BY-4.0. Please credit the original kirtan artists and channels.

Citation

@dataset{gurbani_kirtan_v2,
  title={Gurbani Kirtan Dataset V2},
  author={Surinder Singh},
  year={2026},
  url={https://huggingface.co/datasets/surindersinghssj/gurbani-kirtan-dataset-v2}
}