History
The History page keeps a searchable record of your recent transcriptions and the text TypeWhisper has produced. It also turns the edits you make into correction suggestions that can feed your dictionary.
History stores your transcriptions so you can search and review recent results. The search box matches against the final text, the original raw transcription, and the source app name. An App filter narrows the timeline to entries from a single application (the filter list is built from the apps that appear in your history).
You can export the entries currently shown to CSV, Markdown, JSON, or plain text.
Play caveat: an entry's Play button only works while its session WAV still exists on disk. Because dictation WAVs are deleted on shutdown, this is generally limited to entries from the same session.
Note: History can contain personal or sensitive transcribed text. Treat it like any other record of what you have dictated, and use the retention controls below if you do not want transcriptions stored indefinitely.
Each history entry has an Inspect / view source panel that shows exactly what TypeWhisper produced and what it sent to any LLM for that entry. It has two parts:
- Raw → final diff — a word-level comparison of the raw transcript against the final inserted text, so you can see precisely what cleanup, translation, or a prompt action changed.
- One card per LLM call — for every LLM call behind the entry (a prompt action, cleanup, translation, or memory extraction), a card showing the stage, the provider and model, a local-vs-cloud badge (whether the text stayed on this machine or went to a cloud provider), the exact system and user prompts, any injected memory context, and the model's response.
This makes it auditable what left your machine for each dictation, rather than having to trust that cleanup "just ran locally."
Opt-in. Capturing the full prompt and injected memory context is more sensitive than storing the plain transcript, so it is off by default. Turn on Capture LLM prompt provenance on the Advanced settings page to record it for new dictations. While it is off, the Inspect panel says so (and points you to Advanced) instead of looking as if no LLM was used. It also requires history saving to be on, since it attaches to the stored entry.
Spoken commands (keyphrase mode) are recorded here like dictations, tagged with a violet "Command" badge in the entry header so they're recognizable at a glance. The entry preview shows the result; the raw → final diff reads source text → result (the selected text for an edit, the command itself for a create), and the command's LLM request/response appears in the Inspect panel when provenance capture is on. Only commands that actually landed are stored — cancelled or failed commands are not.
History retention is configured on the Advanced page, with modes Duration (default 90 days), Forever, or Until app closes, plus a Save to history toggle for runs you do not want stored. See Advanced settings for the full set of retention controls.
TypeWhisper generates correction suggestions from the edits you make to entries in history. These suggestions can optionally be auto-learned into your dictionary, and they carry confidence scoring so you can judge how reliable a suggested correction is.
To have corrections saved to your dictionary automatically, you must turn on Auto-learn approved history corrections on the Dictation settings page — it is off by default. With it on, clear phrase-level corrections from your history edits flow into the Dictionary as entries (tracked with a CorrectionSuggestion or AutoLearned source) so future transcriptions apply the correction automatically.
With it off, your edits still generate suggestions, but they are surfaced inline on the edited entry for review instead of being saved silently. Each suggested correction appears as a row with editable Original and Replacement fields, a checkbox to approve it, and a confidence label. Use Save approved corrections to write the checked rows into the Dictionary, or Dismiss to discard them.
Once a correction is learned, it appears in the Dictionary with a CorrectionSuggestion source (Correction is the entry type, not its source) — see the entry screenshot on the Dictionary page.
- Advanced settings — history retention modes, the save-to-history toggle, and the LLM-provenance capture setting.
- Dictionary — where auto-learned corrections land and how dictionary entries are tracked.
- Prompts & AI Actions — spoken commands, which are recorded here with a Command badge.
| Date | Change |
|---|---|
| 2026-06-17 | Initial version. |
| 2026-06-17 | Added History list and entry-detail screenshots. |
| 2026-06-17 | Documented the Auto-learn approved history corrections requirement. |
| 2026-06-17 | Fixed the learned-correction source (CorrectionSuggestion, not Correction); added the App filter, search scope, export formats, Play/session-WAV caveat, and the inline correction-review UI shown when auto-learn is off. |
| 2026-06-17 | Removed unused screenshot placeholders. |
| 2026-07-08 | v0.12.0: documented the Inspect / view-source panel (raw→final diff, per-call provenance with local-vs-cloud badge and response) and its opt-in capture setting, plus spoken commands appearing here with a Command badge. |