Changelog
What changed, and when.
A dated record of what shipped in templobe. Nothing is listed here until it is in the build, and entries stay even when a later result is less flattering than the first one. If a claim here turns out to be wrong, it gets corrected rather than deleted.
November 2026
2026-11-03 — Nine search-facing pages, one generated sitemap
Seven topic pages went up, each answering one question: bulk transcription, offline transcription, research interviews, podcast episodes, free transcription on a Mac, batch limits, and how the options compare. A guides hub and a benchmark page sit above them. A page cannot go live without appearing in the sitemap, because both come from the same list of pages. Every URL in the sitemap at that point, eleven of them, was submitted to IndexNow after deployment.
The benchmark page publishes the numbers from the October 22 run, including the weak one. A separate answer block was added to the homepage because a search result should be able to quote a direct answer instead of a slogan.
2026-11-02 — Website relaunched in English
The site had been bilingual. It is now English-only, matching an English-only beta. Old /zh/ links redirect to the homepage rather than returning 404. The privacy policy was published the same day, and the beta reservation form opened.
The privacy page states plainly that the desktop app is not distributed yet, rather than implying a download and explaining the gap in small print.
October 2026
2026-10-22 — English accuracy measured on five public datasets
FLEURS scored 5.93% word error rate over 194 clips, LibriSpeech test-clean 1.68% and test-other 3.23%, VoxPopuli 6.16%, and accented English from EDACC 19.18%. The accented set is the known weak spot and stays on the page. The long-audio result covers two clips only, so it is published as indicative rather than conclusive.
The same run compared several alternative English configurations. None beat the default overall, so the default stayed.
2026-10-24 — Retranscribing no longer looks frozen
Choosing a different model for an existing file used to change nothing on screen until the new transcript appeared. The detail view now creates a pending entry immediately, marks it as transcribing, and disables the action until it finishes or fails.
2026-10-20 — End-to-end tests against the real app
Thirteen automated end-to-end tests now drive the real application: import, single-file transcription, the batch queue, long-audio handling, model management, the detail view and exports. The suite catches regressions in the screens a person touches, not in isolated functions.
2026-10-18 — Text formatting
Punctuation, paragraph breaks that do not split a word, and conservative number and date normalisation. The formatting step is deliberately cautious: rewriting a transcript aggressively makes it read better and makes it less accurate.
2026-10-16 — Automatic model selection
The app now chooses the right setup for each file instead of treating a mixed folder as one language. The change was measured before it was wired in, and single-language recordings improved.
2026-10-14 — Batch queue, exports, model switching
Batch import and a queue, export to txt, md, srt and vtt, file deletion with a recycle bin, a model manager that can add or switch models, and a home screen that shows what is running. This is the release where the app stopped being a single-file demo.
How this log works
An entry is added when a change reaches the build, not when it is planned. Dates are taken from the requirement records written while the work was accepted, so a date here is the date it shipped rather than the date this page was edited.
Changes that are later reverted stay visible, with a note saying so. A changelog that only contains successes is a marketing page wearing a changelog's clothes.