Words from kirana_natural_word_order_500.json.
Create a session → record single words or full phrases (with optional phrase-line preset) → build audio augments.
Dataset folder: yourname_YYYYMMDD_HHMMSS.
0. Assignments
Sign in with your Gmail account to see and start your assignments.
Google sign-in is not configured on this server (set GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET).
Signed in as
Pick an assignment to start recording.Each take is saved to the server as you record — admin merges from there.
Use this after admin changes — the page does not auto-poll.
Loading assignments…
Progress
Loading…—
Hold to record → release to hear playback → click Sounds good → Next when ready.
1. Record one sentence at a time
Start an assignment above, then record each line below.
Hold record — line appears in the box below
—
—
Playback appears here after you record
Start an assignment to see the recording queue.
All lines in batch
1b. Session recordings
Preview WAVs recorded in the active session only.
No recordings in this session yet.
2. Build augmented dataset
Uses only sentences you recorded. Each line gets 8 audio augments.
Val = every orig (clean) take per line;
train = all augmented copies (gain, noise, speed, …).
Manifest text is the full Meetei line you spoke.
Scaling to a few thousand samples
Record 2–3 takes per phrase (re-record overwrites — duplicate phrase rows for multiple takes).
Enable NeMo speed perturb during training (already on in job.yaml).
Mix this dataset with a larger public Indic/NE corpus in DigitalOcean Spaces.
Add custom noise files under data/noise/ (future: room-specific clips).
TTS from Meetei Mayek text if you have a voice model — bulk synthetic lines.