Skip to main content
Rekody can save your dictations as training data. Review is the page where you clean that dataset: play each clip, fix its transcript, and end up with labels you trust. Open it from the Mac app under Settings → Dictation → Review your dictation data, or from the terminal:
The page runs on your Mac at http://127.0.0.1:7878. Nothing about your audio or transcripts leaves the machine.

​
The second opinion

On first run you choose how to review:
  • Get a second opinion (recommended). Whisper large-v3 re-listens to every clip and flags where it disagrees with your saved transcript, so you review only the differences. It is a one-time 3.1 GB download from Rekody’s model repository, checksum verified, and it runs entirely on your Mac.
  • Start from my transcripts. No download. You read and fix each transcript yourself, and you can add the second opinion later.

​
Reviewing a clip

Red words are your original, teal words are the suggested fix. Click any word to switch which one gets saved, then:
  • Save correction (S) saves exactly what the WILL SAVE line shows.
  • Keep original (K) marks the clip reviewed with no changes.
  • Edit text (E) opens the transcript for free editing.
  • Delete clip removes the audio and transcript from your dataset. It always asks first, and it always takes a click: deletion has no keyboard shortcut and cannot be undone.
Use the two dropdowns to filter by status and by how much the suggestion differs.

​
Reviewing from your phone

Start the server with the network flag and open the printed address on a phone connected to the same Wi-Fi:
Treat this as a trusted-network feature: while the server runs with --lan, anyone on your network can open, edit, and delete the dataset. The default (no flag) stays private to your Mac.

​
How much you have reviewed

The top bar leads with hours: “3.2 hours reviewed · 41 awaiting review”. The hours count the audio behind clips you have decided on (kept, corrected, or edited); deleted clips are left out, since their audio is gone. Under an hour it reads in minutes (“48 minutes reviewed”), and from an hour up it reads with one decimal. Only reviewed audio is counted. There is no remaining-hours number, because the dataset keeps growing every time you dictate, so anything measured against it would move under your feet.

​
When you finish

When nothing is awaiting review, the done screen says plainly what you now have (clips reviewed, hours of audio) and hands it to you: download your reviewed clips as one zip, write a cut folder into your dataset, or bring in a cut from another Mac. Rekody does not train a model for you. The export is your own data, for whatever you want to do with it. If you want to hear from us when there is something to try, there is a Request early access link on that screen.

​
Export reviewed clips

An export is a cut: a training-ready set holding only the clips you reviewed (kept, corrected, or edited). Deleted clips and clips still awaiting review stay out. On the done screen, pick a date (clips from that day or earlier are included; the default, today, includes everything) and press Download the zip. Your browser saves one cut-<date>-<hash>.zip to Downloads, audio included, so the file is complete on its own. Nothing is uploaded anywhere: the zip is built on your Mac and handed to the browser that asked for it. If you would rather have a folder inside your dataset, the same card has Write a folder into my dataset instead, with a Copy audio files tickbox. The result line shows the clip count, total audio, and the absolute path of the cut. Each cut lands in exports/ inside your dataset, named cut-<date>-<hash>, and contains:
  • manifest.jsonl, one line per clip: audio_filepath, duration, engine, text (the label as you reviewed it), label_source (original, teacher, or human), and decided (when you reviewed it).
  • cut.json: created, until, dataset_root, clip_count, total_duration_secs, and manifest_sha256 (the checksum of the cut’s manifest).
Without Copy audio files, the manifest references the audio in your dataset. With it, the flac files are copied into the cut and the paths point inside the folder, so you can hand the whole thing to a training job.

​
Export from the terminal

The same export runs headless, no server needed:
  • --until YYYY-MM-DD includes clips dated that day or earlier (default: today).
  • --copy-audio copies the audio files into the cut so the folder is self-contained.
  • --out DIR writes the cut into DIR instead of exports/ in the dataset.
  • --zip produces one cut-<date>-<hash>.zip instead of a folder. A zip always carries its audio, so it is complete on its own; there is no unzipped copy left behind.
Agents can drive this. On success the command prints exactly one line to stdout, everything else goes to stderr:
That mirrors the REVIEW_URL= contract the server prints, so a coding agent or script can run your export and read the resulting path with one line. The path points at whatever was produced: the folder, or the zip when you passed --zip.

​
Move a cut to your other machine

Review on your work Mac, review on your personal Mac, and merge the two. Export a zip on one machine, get it to the other however you like (mail it to yourself, drop it in a shared folder), then import it:
  • In the page: Import a cut in the toolbar, or Choose a cut to import on the done screen. Pick the zip and the merge runs on the spot.
  • In the terminal: rekody review import ~/Downloads/cut-2026-08-19-1a2b3c4d.zip. A cut folder works too.
What a merge does, clip by clip:
  • Clips are matched by their audio file name. Audio you do not already have is copied in; audio you have is left alone.
  • Each incoming correction is replayed as if you had made it here, so the transcript, its corrected flag, and its label source all land the same way.
  • The newer review wins. If both machines reviewed the same clip, the one decided later stays and the other is skipped and counted. Nothing is silently overwritten.
  • A clip you deleted on this machine is never brought back. Deleting is meant to be final.
  • Merged decisions are stamped with the cut they came from, so you can see where a label originated and undo still works.
  • manifest.jsonl and decisions.jsonl are copied to .bak-import files before anything is written.
Importing the same cut twice does nothing the second time, so it is safe to re-run if you lose track. The summary tells you what happened: clips merged, audio files copied, clips left alone, and your new reviewed total. rekody review import prints one stdout line for scripts and agents: