WSE Transcribe

v1.18.2 2026-10-27

Fix: Release notes and confirmation dialogs keep keyboard focus inside, close with Escape and return focus to their opener. Closed release notes no longer interrupt keyboard navigation.
Fix: Audio seeking and every transcript timestamp work with the keyboard. Tab reaches speaker names, timestamps, text and the Save/Revert actions without trapping focus or discarding edits.
Improvement: Speaker names and important status text are easier to read in both themes. Mobile navigation and controls have larger targets, and expanded Nextcloud help fits narrow screens.
Improvement: Interface styles are prebuilt and shared scripts are served locally at fixed versions, avoiding browser-side compilation and floating JavaScript dependencies.
Security: Media parsing now limits input formats, protocols, time and resources, and cleans up child processes after completion or interruption. The runtime and its dependencies have also been updated. These safeguards reduce exposure; they do not eliminate native-parser vulnerabilities.

v1.18.1 2026-10-13

Fix: Assigning several voices to the same participant now combines them into one speaker in the legend, colours and recording counts. Previously saved assignments work after reload; Undo and Reset still restore the original voices.

v1.18.0 2026-10-13

Feature: Optional participant roster for Identify speakers, reusable assignments, global name correction, Undo and confirmed Reset. Original speaker labels are preserved.

v1.17.2 2026-10-09

Fix: Very long meetings (tested with three hours) are transcribed reliably. They are split into as few parts as possible, at most about 90 minutes each: a three-hour meeting becomes two parts with one join instead of five parts with four. If the transcription service runs out of time on a part, the app tries again with shorter parts, never shorter than 30 minutes.
Improvement: Long recordings are prepared as compact speech audio, so each part keeps good quality within the upload limit. Stereo is mixed to mono only when both channels carry the same sound or one is silent.
Improvement: A long meeting reads as one continuous transcript. Parts are joined in a pause, so no word is cut, and a speaker's turn simply continues across the join. Downloads and the Nextcloud copy contain no join markers. The editor marks a join only where a speaker could not be matched across it; if one person then has two labels, merge them.
Fix: Short outages of the transcription service (timeouts, connection errors, HTTP 500/502/503/504) are retried automatically instead of failing a long job.

v1.17.1 2026-10-09

Fix: Long meetings that exceed the transcription upload limit after compression are split automatically into large, overlapping parts. The original recording and playback timeline stay intact. Speaker labels are matched across shared speech; check labels at marked stitch points, especially when someone was silent there. Smaller recordings still use one shared speaker context.
Improvement: Removed the decorative waveform from the recording drop zone on the start page.

v1.17.0 2026-10-08

Feature: A redesigned start page. Signed out, it says who signs in where and shows a short example transcript. Signed in, a chosen file shows its size and length, you can drop a recording anywhere on the page, the Transcribe button shows the upload progress, and your three latest transcriptions are listed below.
Feature: A new waiting screen while a transcription is queued or running. Pick its animation in Settings: Anemone (default), Mycelium, Stream or Tree rings. The choice is saved to your account.
Feature: Running transcriptions show the time left, estimated from the recording length and how long recent jobs took. Until enough longer recordings have finished, the waiting screen shows elapsed time only. If a job runs over, it says so instead of stalling at the end.
Fix: Upload errors such as an unsupported file type, a file over 2 GB or the hourly upload limit now appear in the page instead of as raw text.
Fix: A finished transcription sends one browser notification instead of two. The job page polled the server twice while waiting, and opening an already finished job no longer asks for notification permission.
Improvement: Secondary text has more contrast in light and dark mode.

v1.16.1 2026-10-08

Feature: 🎉 Our best transcription model is now available to everyone: ElevenLabs Scribe v2! 🥳 No personal API key needed. Just upload or record.
Fix: Meetings are transcribed as one complete recording, with one shared speaker context. No artificial stitch points or new speakers at minute boundaries. Large audio is compressed without shortening pauses or changing playback speed; the original stays intact.

v1.16.0 2026-10-08

Feature: Transcriptions now use ElevenLabs Scribe v2 with centrally managed access. Uploads, browser recordings and retries work without a personal API key.
Improvement: Upload and recording controls are simplified. Audio keeps its original speed, pauses and playback timeline.
Note: Review speaker labels before exporting, particularly at chunk boundaries. Speaker names and merges remain editable.

v1.15.3 2026-10-07

Fix: Clicking anywhere on a speaker label focuses its rename editor and selects the existing name immediately. Later clicks in the active editor still place the caret normally.

v1.15.2 2026-10-05

Fix: Nextcloud recordings and transcripts from meetings with the same date and title now use separate job folders instead of overwriting each other. Folder names include the full job ID within the UTF-8 byte limit; renamed meetings and updates keep their previously stored folder.
Fix: Small Mistral chunks now retain the transcript tail and correct timestamps. Byte-limit splitting and recursive 413 retries carry their rendered schedule into overlap trimming and speaker reconciliation, so the configured overlap no longer discards unique content when the splitter reduces it.
Fix: Nextcloud connection failures and timeouts no longer turn a saved speaker rename, merge, or transcript edit into an HTTP 500. The remote sync error is recorded separately for retry; manual Store/Update returns its normal failure response without exposing transport details.
Fix: Markdown and text exports work with Unicode meeting titles and original filenames. Download names include a UTF-8 filename and an ASCII fallback; quotes and control characters cannot corrupt the response header. Transcript content and stored titles stay unchanged.
Fix: Nextcloud recording uploads stream in 64-KiB chunks with an explicit content length, so storing a large recording no longer buffers the entire audio file in a web process or worker. Transcript uploads and storage modes remain supported.

v1.15.1 2026-09-02

Fix: Opening an installed PWA from an older anonymous page no longer starts a redundant OAuth flow when the app already has a valid session. GitHub and Google sign-in now return directly to the app in that state instead of risking an "Invalid OAuth state" page that disappears after a reload.
Fix: Restored or replayed OAuth callback pages are now harmless after sign-in has already completed. They return to the app without exchanging the authorization code again, and a callback without a code produces a controlled authentication error for signed-out users instead of FastAPI's generic validation response.
Fix: Renaming a speaker no longer storms the server. With the guided panel open and a save failing, clicking between two gutter labels used to trade focus and re-fire each other's save endlessly (over a thousand requests in three seconds); only a reload stopped it. The failure path no longer pulls the cursor back when you have moved to another field, so the loop cannot form.
Fix: Clicking through speaker labels without changing anything no longer re-saves and re-uploads. A focus/blur that leaves the name untouched now skips the request entirely — on the client, and short-circuited on the server too — so merely reviewing a transcript stops re-uploading it to your Nextcloud once per label touched.
Fix: A rename the server did not confirm is never treated as already saved. If a save fails, leaving the field again retries it instead of skipping it as unchanged, so a name can no longer sit on screen looking saved while the server has never seen it. This holds whichever occurrence of the speaker you use: a speaker with several turns has a name label on each of them, and they now share one record of what the server acknowledged. Previously each label kept its own, so saving from one turn and then editing from another could skip the request and leave the stored name differing from the one shown everywhere on screen.
Note: Because a redundant save no longer reaches the server, a Nextcloud upload that failed during a rename is no longer retried by clicking through the same label again. Use "Store/Update" on the job page to repair it.

v1.15.0 2026-08-17

Fix: The guided speaker annotation now has its own name field. Previously the panel showed the speaker label as plain text that looked editable but was not, and the field it actually focused was the one in the transcript, far below the visible area with no cursor in sight. Typing worked, blindly, and nothing on screen said so. The name goes in the panel now, right under the audio sample, and the transcript no longer scrolls away while you work.
Fix: "Another sample" is audio only again. With the cursor still in a speaker label it could save a half-typed name and skip that speaker.
Fix: Enter in the panel keeps working after you edit a segment's text. Editing the transcript around the speaker being named used to leave Enter doing nothing at all, with no hint why.
Fix: Escape no longer leaves the "What's new" window open when you use it to leave the annotation pass.
Improvement: Pressing Enter on an empty name field resets that speaker to its placeholder and stays put, instead of saving an empty name. Unrecognised voices keep their place in the count.

v1.14.0 2026-08-12

Feature: Rescue a failed transcription. The failure card now offers "Download recording", so your audio is never trapped behind a broken API key, and Retry lets you pick the transcription engine instead of repeating the one that just failed. ElevenLabs is preselected only when a key is stored for your account; otherwise Mistral/Voxtral is, with a direct link to add an ElevenLabs key.
Feature: "Save a copy to this device" in the recorder. A browser recording can be saved locally straight from the preview, before it ever leaves your machine.
Improvement: Failed jobs keep their audio for 72 hours (`FAILED_AUDIO_RETENTION_HOURS`), counted from the failure, so a job that fails overnight is still recoverable the next working day. Finished jobs keep the existing 24 hours.
Improvement: ElevenLabs errors now say what to do about them. Pasting a key ID instead of an API key is named as exactly that (real keys start with `sk_`), both when saving the key in Settings and on a failed job; an exhausted quota is reported as a quota problem rather than a rejected key; and the raw response headers no longer land on the job page. Settings also shows when your stored key was last validated.

v1.13.0 2026-08-07

Feature: Identify unnamed speakers in one guided, keyboard-first pass. WSE Transcribe selects a substantial early sample for each speaker, plays only that segment, focuses the existing inline name field, and advances after Enter. Tab and Shift+Tab navigate without saving, Escape cancels, and Replay or Another sample help when the first voice sample is unclear.
Improvement: Guided annotation stays synchronized when a name is cleared or two speaker labels auto-merge. Failed saves keep the draft and current speaker in place, completion is explicit, and the mode appears only when playable audio and unnamed speakers are available.
Fix: Speaker labels now use natural numeric ordering everywhere, including transcript legends and merge controls. `speaker_10` follows `speaker_9` instead of appearing between `speaker_1` and `speaker_2`; malformed labels still have deterministic fallback ordering.

v1.12.0 2026-07-22

Feature: Store recordings and transcripts in your own Nextcloud. Add a WebDAV connection in Settings (URL, username, and a Nextcloud app password) — saving runs a real write test so a saved connection is always known-good. Choose what to store: recording and transcript (default), recording only, or transcript only.
Feature: "Transcribe and store in Nextcloud share." Tick the option when uploading and, the moment the transcription finishes, the result is saved to your Nextcloud automatically. A storage problem never fails the transcription — the error is shown on the job so you can retry.
Feature: "Store in Nextcloud share" / "Update in Nextcloud share" buttons on each finished transcript. The button appears once the transcript exists; after you edit speakers or text it switches to "Update" so you can push the corrected version. Each job lands in its own dated folder (for example `WSE-Transcribe/2026-07-09_Team-meeting/`) with the recording and a `transcript.md`. Success or failure is reported inline.
Feature: The "Transcribe and store in Nextcloud share" option now also appears in the Record-audio tab, and is ticked by default once you have a Nextcloud connection. Before it is shown, your share is checked live: if it does not respond, the option is still offered but highlighted in a warning colour and left unticked so nothing silently fails.
Feature: Automatic Nextcloud sync on edits. Once a transcript is stored in your Nextcloud share, changing it (renaming speakers, editing text, merging or unmerging) re-uploads the updated `transcript.md` to the same folder automatically, and a toast confirms the share copy was refreshed. Jobs you never stored are left untouched, and "Recording only" storage is unaffected.
Improvement: Settings polish — the intro now describes all sections; saving any setting keeps you on the same block instead of jumping to the top; the "What to store" selector spans the full column width and confirms with a "Storage choice saved" message; and "Test write access" now sits next to "Save & test write access".

v1.11.0 2026-07-01

Feature: Push notifications. You now get a browser push even when the app is closed: when someone shares a transcript with you, and when your own transcription finishes. Turn it on with the existing notifications switch in Settings; your device subscribes and delivery happens in the background. Works on Chrome, Edge, Firefox and installed PWAs; other browsers keep the in-app notification.
Improvement: The "finished" notification now also arrives when the tab is closed (previously only while the app was open), and the foreground and background versions collapse into one so it never shows twice.

v1.10.0 2026-06-27

Feature: Desktop notification when a transcription finishes. Your browser notifies you the moment a transcript is ready — even if the tab is in the background. On by default; switch it off anytime in Settings.
Feature: Your theme preference now follows your account. The light/dark choice is saved to your settings and applied automatically when you sign in, on any device — with no flash of the wrong theme on load.
Feature: "What's new" opens automatically after an update. When you sign in to a new version, the release notes appear once so you never miss what changed.
Improvement: New Preferences section in Settings to manage notifications and theme.

v1.9.0 2026-06-27

Feature: Installable app (PWA). WSE Transcribe can now be installed to your phone or desktop straight from the browser and launched like a native app — its own icon, a standalone window, and a name on your home screen or taskbar.
Feature: Offline fallback page. Opening the app without a connection now shows a friendly "You're offline" screen instead of the browser's error page. (Uploading audio and loading transcripts still needs a connection.)
Improvement: App icons and a matching theme color for installs, including maskable icons for Android adaptive icons and an Apple touch icon for iOS.
Improvement: Privacy-preserving caching. A conservative service worker caches only the app shell (icons and the offline page) and leaves everything private — logins, jobs, transcripts, and audio — untouched, so nothing sensitive is stored on your device.
Improvement: Works behind a sub-path. The manifest and service worker are `BASE_PATH`-aware, so installs work whether the app is served at the root or behind a reverse-proxy sub-path.

v1.8.2 2026-07-01

Fix: Long recordings transcribed with Mistral/Voxtral no longer lose speech at chunk boundaries. When a speaker was talking across a chunk boundary, the overlap-head trim dropped that entire turn (it starts inside the overlap head), discarding all of its post-cutoff content — on some recordings roughly half the transcript. Boundary-spanning turns are now kept and clipped to the cutoff, with the duplicated overlap words removed proportionally by duration. Measured on a real interview clip, transcript coverage rose from ~51% to ~98% of the whole-file reference with negligible duplication.

v1.8.1 2026-07-01

Fix: Speaker names no longer shift at chunk boundaries ("stitch points") in long recordings transcribed with Mistral/Voxtral. Diarization speaker IDs are only consistent within a single API request, so when a long file is split into overlapping chunks the same person could be renumbered between chunks and inherit the wrong name after a boundary. Chunks are now reconciled by matching speakers that overlap in time inside the shared window, so a speaker keeps one label (and one assigned name) across the whole transcript. A speaker who cannot be verified in the overlap (silent across the boundary) is given a fresh label rather than risk inheriting another speaker's identity; this may over-segment a speaker into extra labels, which you consolidate with the existing merge UI — naming two labels identically auto-merges them. The result never silently mislabels a speaker; a larger overlap (`MISTRAL_CHUNK_OVERLAP_SECONDS`) reduces over-segmentation. ElevenLabs transcriptions were unaffected (whole file, single request).

v1.8.0 2026-05-14

Feature: Reversible speaker merges — every merge can now be undone via a one-click "Undo" button in the transcript toolbar. Merge history is stored in the database and persists across page reloads.
Feature: Drag-and-drop speaker merge — drag one speaker's name chip onto another in the toolbar legend to merge them instantly. Visual feedback highlights the drop target.
Feature: Auto-merge on identical names — renaming a speaker to exactly match another speaker's custom name triggers an automatic merge (also undoable).
Improvement: The merge panel now stays open after performing a merge, so you can merge multiple speakers in a row without reopening it.
Improvement: Toast notifications for merge, undo, and auto-merge actions.

v1.7.3 2026-04-30

Fix: Browser recordings whose WebM duration is reported as `N/A` no longer fail with `could not convert string to float: 'N/A'`. Duration detection now ignores non-numeric ffprobe values and falls back to packet timestamps.
Fix: Job timestamps are rendered in the app timezone (`Europe/Berlin` by default) instead of showing raw UTC.

v1.7.2 2026-04-30

Fix: Browser recordings (MediaRecorder `.webm`) failed with `Unexpected error: 'duration'` because the WebM container header lacks a duration tag. `ffprobe` now falls back to stream-level duration and packet scanning when the format-level duration is missing.

v1.7.1 2026-04-30

Fix: Audio optimizer silence removal was cutting natural speech pauses at 0.5 s instead of the intended 0.7 s. The v1.6.2 fix raised `DEFAULT_SILENCE_DUR` in the module but `tasks.py` still hardcoded 0.5 s, so the fix never reached production. Now uses module defaults so future tuning takes effect everywhere.

v1.7.0 2026-04-30

Feature: Google OAuth login for Netresearch employees. Sign in with your `@netresearch.de` Google Workspace account alongside the existing GitHub login for WSE-research members.
Feature: Tenant-isolated sharing. Users from different organisations (WSE-research via GitHub, Netresearch via Google) are fully separated: the sharing UI only shows same-team members, and cross-team sharing is blocked server-side.
Feature: Display names from Google profile / GitHub are now shown in the navbar, share panels, and shared-with badges instead of raw login identifiers.
Security: Google `hd` domain claim is verified server-side (the `hd` auth URL parameter is only a UI hint). Unverified email addresses are rejected. Separate CSRF state cookies for GitHub and Google flows prevent cross-flow collision.

v1.6.7 2026-04-29

Fix: ElevenLabs scoped API keys were incorrectly rejected during validation. ElevenLabs now uses permission-scoped keys, and keys without the `user_read` scope caused a 400 error even though they work for transcription. Validation now accepts any key that ElevenLabs recognizes, regardless of scopes.

v1.6.6 2026-04-24

Fix: ElevenLabs could be selected as transcription engine without a personal API key configured, leading to a failed transcription. The option is now grayed out with an "add key" link to Settings until the user stores their own key.

v1.6.5 2026-04-24

UI: added a "What's new" button in the footer that opens a modal with the full version history. Each release is shown with its version tag, date, and change items, styled to match the rest of the app. Closes with Escape, backdrop click, or the close button.

v1.6.4 2026-04-23

Feature: transcript pages now offer owners a download action for the original uploaded audio while the 24-hour audio retention window is active.
Backend: optimized jobs keep two separate files: `audio_path` remains the optimized playback source so transcript timestamps stay aligned, while `original_audio_path` preserves the untouched upload for download. Retention cleanup and delete flows clear both paths together.

v1.6.3 2026-04-23

UI (mobile): transcript segments now switch from the desktop two-column timeline to a compact stacked layout on phones. Speaker + timestamp sit in a small header and the transcript text gets the full content width, so long transcripts are readable without the speaker gutter consuming half the screen.
UI (mobile): the transcriptions overview now collapses the large date gutter into a single metadata line above each ledger entry. Titles, status, provider, duration, speaker count, and word count get the usable card width on narrow screens.
UI (mobile): transcript toolbar/actions, audio controls, and the signed-in navbar now wrap/tighten at the same breakpoint to avoid horizontal crowding while leaving the desktop layout unchanged.

v1.6.2 2026-04-23

Fix (audio optimization): the pre-transcription filter chain was ASR-correct but unpleasant to listen to — which matters because v1.5.2 promotes the optimized file to be the canonical playback source. Four concrete problems fixed:
Also softened the compressor from 4:1 / +6 dB makeup to 2.5:1 / +3 dB — the old setting turned every breath into a roar once NR artifacts were riding on top.
Speed-up (1.4×) and silence removal are kept deliberately: the transcript timeline is aligned with the optimized audio, so the player scroll-sync depends on both.
`scripts/audio_optimize.py` is now a thin CLI over `app.audio_optimizer.optimize()` — the filter chain lives in one place. Unit tests pin the shape of the chain (`tests/unit/test_audio_optimizer.py`).

v1.6.1 2026-04-18

Fix: the active-segment highlight during audio playback in v1.6.0 painted two disconnected rectangles because the 1.75rem grid `column-gap` between the speaker gutter and the transcript body stayed at the page background while both sides were tinted. Removed `column-gap` and moved that spacing into the body's `padding-left`, so the highlight paints as one continuous band. The gutter's cold `border-right` is also softened to an accent tint when the segment is active, so no divider cuts through the warm stripe. The spine dot's outer ring picks up the active bg too, preventing a bright halo.

v1.6.0 2026-04-18

UI (transcript): the reading and editing surface is redesigned as a margin-gutter layout — the speaker name (Fraunces italic 600) + timestamp + colored spine dot sit in a left column, the transcript body flows in the right column. Segments are CSS-grid children of a single 2-column parent via `display: contents`, so all gutters align perfectly. Replaces the previous "speaker badge above a boxed card" layout.
UI (transcript editing): segment text is now contenteditable in place. Clicking any paragraph and typing triggers a yellow highlighter wash and a per-segment Save / Revert toolbar with a pulsing "unsaved" indicator. Keyboard: `Cmd/Ctrl+Enter` saves, `Esc` reverts, `Tab` jumps to the next segment. Save confirmations use a global bottom-center toast that survives HTMX swaps. Replaces the previous double-click → textarea swap.
UI (speaker identification): the separate "Identify Speakers" form card is removed. Speaker names in the transcript gutter are themselves contenteditable. Renaming any instance of a given diarization label live-propagates to every other segment with that label and to the legend chips, then persists on blur via a new `POST /jobs/{job_id}/speakers/{label}` endpoint (returns 204 — the DOM is already up-to-date). The compact toolbar above the transcript now carries only export links, merge-speakers, and a row of colored speaker chips that double as jump-to-speaker anchors.
UI (processing state): the 9 pulsing bars + "this may take a few minutes" are replaced with a vertical staged stepper — Uploaded → Queued → (Optimizing, if enabled) → Transcribing & diarizing → Ready. The active stage has a rippling indigo ring and an indeterminate shimmer; no fake progress percentages since the backend doesn't track sub-stages. Elapsed timer ticks client-side and resyncs on every HTMX poll.
UI (jobs list): each row is now a ledger-style entry: date-first left gutter (Fraunces day number + weekday + time), title, monochrome status pill, provider + filename subtitle, and a metadata row with duration, colored speaker dots + speaker count, and word count. Shared-by shown as a small mono chip inline; shared-with avatars stack on the right. Wrapped in a bright `.ledger` card container that lifts it off the page background. A single `ledger_row` Jinja macro renders Mine / Shared / Archived tabs; batch-select and batch actions (Archive, Delete, Share) are unchanged.
Backend: new `POST /jobs/{job_id}/speakers/{label}` upserts a single speaker's display name and returns 204. New `_job_stats()` helper runs one bulk `Segment` query across all visible jobs to produce `{duration_s, speaker_labels, word_count}` per id, so the jobs list picks up its new metadata without adding an N+1. The `<html>` element now carries `data-base-path="{{ base_path }}"` so client-side `fetch()` calls build correct URLs under `BASE_PATH` deployments.
Repo: design exploration preserved under `design-previews/` — four speaker-coding drafts × three editing drafts with tradeoff notes, an interactive combined prototype, and five further drafts (processing / jobs list / speaker ID).

v1.5.4 2026-04-17

Docs: upload-page subtitle still said "keep the source audio for 2 hours" — the retention was bumped to 24 h in v1.5.0 and the audio is now also used for in-app playback, not just retry. Rephrased to "Jobs keep their audio for 24 hours so you can replay it in sync with the transcript or retry if something fails."
Docs: the "Optimize audio before transcribing" toggle now also mentions that playback plays the optimized version — the timeline is intentionally aligned with the transcript (v1.5.2 behavior).

v1.5.3 2026-04-17

UI: the audio player card now hides itself if the underlying file is gone (e.g. retention cleanup removed the file while the page was still open, or the file was deleted server-side). The `<audio>` element listens for `error` / `stalled` + `NETWORK_NO_SOURCE` and flips an `errored` flag that drives `x-show`. Transcript segments stay visible — only the dead player disappears. Normal retention still nulls `audio_path` server-side on reload, so most users never hit this path; this covers the live-tab edge case.

v1.5.2 2026-04-16

Fix: when a job was transcribed with `optimize_audio=True`, the playback stream was the *original* upload while the segment timestamps referred to the *optimized* timeline (1.4× speed + silence removed). The transcript scrolled noticeably faster than the audio. The optimized file is now promoted to be the canonical playback source on success — `job.audio_path` points at the optimized file and the original is removed (a `done` job has no retry path that would need the original). On a failed run the optimized leftover is cleaned and the original is retained as before, so retry still works.

v1.5.1 2026-04-16

Fix: transcript segments arriving via HTMX swap (i.e. when a fresh upload finishes) showed empty text boxes. Root cause: the Alpine `segmentEditor` / `transcriptPlayer` / `audioPlayer` definitions were inside the swapped fragment, so `Alpine.data()` registration ran *after* `alpine:init` had already fired and the new `x-data` divs were processed. Moved all Alpine component registrations and player CSS into `base.html`, where they load once at initial page load — every subsequent HTMX swap now binds correctly. Also dropped the `x-text="text"` indirection that wiped server-rendered content if Alpine init failed.
UI: replaced the native `<audio controls>` (which rendered as a jarring white pill in dark mode) with a custom player themed to the app's surface/accent palette in both light and dark mode. Play/pause, time display, hover-preview seekbar, and mute toggle.

v1.5.0 2026-04-16

Inline transcript editing: double-click any segment to correct the text. Changes are persisted to SQLite and reflected in `.txt` / `.md` exports. Cmd/Ctrl+Enter saves, Esc cancels, clicking away discards the edit. (#9)
In-app audio playback with timeline sync: the original upload now streams from a new Range-capable `/jobs/{id}/audio` endpoint. A compact `<audio>` player above the transcript auto-highlights the currently-playing segment, the viewport follows playback unless you've scrolled away, and clicking any segment timestamp seeks the audio. (#10)
Default audio retention raised from 2 hours to 24 hours (`AUDIO_RETENTION_HOURS`) so users can replay or retry for a full day. (#10)

v1.4.1 2026-03-23

Fix Mistral large-upload handling with size-aware chunking and automatic 413 fallback.
Preserve source audio for a short retry window and add an owner-only retry flow for failed jobs.
Prevent active jobs from aging out during processing by basing retention cleanup on terminal `status_updated_at`.
Add conservative chunk-boundary stitching and transcript UI seam markers so split transcripts render more gracefully.
Fix a real-file regression where already-rendered chunks could be re-split recursively instead of being uploaded directly.
Apologies again to Jonas Kunze (jbkze) for the dataloss becuase of the bug. should be fixed now.