Live Captioning for Gurbani Kirtan

A benchmark for systems that identify which line of Gurbani is being sung, in real time.

Baseline visualizations

Each page shows GT vs Pred vs Diff for every second of audio. Play audio and hover to see the canonical Gurmukhi line at each moment.

	Shabad-aware (oracle)	Shabad-unaware (blind)
Offline	easiest — line alignment only	identify Shabad, then align
Live	follow along in real time	hardest — identify & follow live

A prototype for the hardest variant (live + blind) is running at bani.karanbirsingh.com and scores ~65–70% — per-case visualizations in the blog post.

Synthetic baselines
Baseline	Description	Score
empty	No predictions — silent everywhere	26.0%	view →
shifted_5s	Ground truth delayed by 5 seconds	85.5%	view →
perfect	Exact copy of ground truth	100.0%	view →

Get involved

• Test your system — run it against the benchmark and see how it scores. The README has the submission format.
• Annotate more recordings — the benchmark only has 4 Shabads today. More annotated recordings make it more representative. One idea: use a real STTM volunteer session as ground truth.
• Expand the benchmark — multi-Shabad recordings, Katha interludes, Simran, transitions between Shabads.

Get in touch: admin@karanbirsingh.com

Get started on GitHub →