Fix: transcribe to Urdu then transliterate to Roman (no Gurmukhi, no translation)

Fix Roman draft: use translate task instead of auto-detect (was outputting Gurmukhi)

Save every Groq inference to CSV immediately on clip load

Add Roman + Urdu dual Whisper drafts, both saved to CSV

Switch Whisper to Groq free tier (28800 sec/day, no credits)

Fix go_back missing 9th return value (state_whisper_draft)

Add Whisper debug logging to diagnose silent failures

Always save Whisper draft to CSV regardless of volunteer edits

Use Aqiba HF token for Whisper API inference

Run Whisper small locally in Space — no API credits needed

Add back button with undo, Whisper status indicator, history tracking

Sort clips shortest first for faster volunteer experience

Add loading indicator on Start button click

Add Whisper auto-draft via HF Inference API

Fix: use Audio(decode=False) + soundfile to bypass torchcodec

Fix: serve audio server-side, add Urdu translation field, add soundfile dep

Fix: remove audio column before iterating to avoid torchcodec error

Resolve README conflict, use HF frontmatter

Initial Memoni labeling tool

initial commit