Welcome to ASR Workstation
Select a data folder containing paired .wav and .txt files to begin reviewing transcriptions.
Select a data folder containing paired .wav and .txt files to begin reviewing transcriptions.
Shown on the team dashboard when you share progress. Counts only — no audio or transcripts are sent.
A quick guide to labeling efficiently. Reopen it anytime with the ? button next to the app title.
A Workflow toggle in the sidebar lets you choose how you label (your choice is remembered). Either way you work in chunks — set the chunk size in the sidebar (e.g. 500).
If the audio is garbage — noisy, empty, clipped, or you simply can't make out what is said — delete it and move on. Don't replay the same clip three times hunting for words. A clean dataset beats a big one. Deletions are only logged to deleted_segments.json (the files stay on disk), so nothing is truly lost.
The corrected transcripts are what make this dataset worthwhile — do them carefully and make the text match exactly what is spoken. In practice only about 10–15% of files need a real fix; if you find yourself editing far more than that, you're probably over-correcting.
The transcription box has Hindi transliteration on by default. Type a Hindi word in Roman letters and press space — e.g. namaste becomes नमस्ते. When several options exist, a small dropdown lets you pick the right one. Turn it off anytime with the checkbox above the box if a transcript is in plain English.
Space play/pause · Triage: D delete, → already correct, C needs correction · Correct: Ctrl/Cmd+Enter submit, → skip · Shift+←/→ navigate.progress.json file in your data folder, so you can close and resume later, or pick up on another machine by reopening the same folder.If you lost your correct/edited flags but kept your work up to a point, mark the first N files as done and resume right after them. Existing deletes and anything already done are kept.