Hear a short clip from your Nytka recordings and tap what kind of audio it is, so Nytka's models learn from your ear.
Nytka is a self-hosted second brain: a pendant records the day, the server transcribes it and pulls out conversations, people and tasks. Some of what the pendant hears is not a person in the room: a TV, a laptop, the far side of a call on speaker. Some lines are too quiet to trust, and some voices are the owner's own but scored as someone else's. My ear tells these apart in a second; the models need labelled examples to learn the same. Earlier rounds of labelling on throwaway pages already moved real settings, and one such page lost every label to a button bug. This app makes labelling a habit with a test behind every button.
- A vertical feed of about 30 clips, one per screen, each playing as it arrives. The server picks the clips its models are least sure about, plus a random slice so rates stay honest; "Mixed" interleaves tasks in short runs.
- One fixed yes/no question per task: swipe right for yes, left for no, up to skip, down to go back, hold for every answer and the optional details (device type, kind of sound, quality flags, language, a corrected transcript). A button rail does the same for anyone who doesn't swipe; arrow keys and 1–7 on a desktop.
- The model's guess stays hidden until after the answer, so it never steers the ear. Every answer can be undone for 4 seconds. The feed ends at the day's goal with what the session moved, and nothing after it.
- Answers wait in the browser until the server confirms them, so a reload or a dropped connection loses nothing. The feed pauses when the tab is hidden and never plays in the background.
- Lines marked "me" teach the voiceprint only after a separate confirmation, because that cannot be undone.
- Clips and labels stay on the Nytka server. This repository holds code only; its tests use synthetic tones.
The taxonomy, what each label is for, how the server picks clips, and the privacy rules: DESIGN.md. The API both clients use: docs/api.md. The Android client: nytka-app/labeler-android.
Run it against the fake API (synthetic clips, nothing leaves your machine):
npm ci
npm run dev # open http://127.0.0.1:8787/Run the image in front of a Nytka server that has the labeling routes:
docker run -p 8080:8080 \
-e NYTKA_UPSTREAM=http://nytka:8080 \
-e NYTKA_LABELING_TOKEN="$(cat token-file)" \
ghcr.io/nytka-app/labeler:latestPut an SSO proxy in front of it: the container adds the token itself, so whoever reaches it can label.
flowchart LR
B[Browser or phone] -->|SSO| P[SSO proxy]
P --> C[Labeler container: static app + Caddy]
C -->|/api/v1/labeling only, adds token| S[Nytka server]
S --> D[(Postgres: clips, labels)]
The page is plain JavaScript with no build step. Caddy serves it and forwards only /api/v1/labeling/*, replacing
any token the browser sends with its own and dropping the proxy's cookies. Every other server route answers 404
from the container.
Apache-2.0. See LICENSE.