github ttlequals0/Audicle v0.63.4
0.63.4

25 minutes ago

0.63.4

Security

  • Reject a DOCX upload that is too large once uncompressed. The limits are 5,000 entries, 200 MB in total and 50 MB of XML parts, checked before the file is parsed. XML parts are identified from the package content types, and a file that repeats an entry name is refused. A 435 KB file that expanded to 343 MB used to stop the worker.
  • Refuse to rasterize a PDF page or decode an image for OCR above 50,000,000 pixels. A page that declared a very large size used to request a bitmap of several gigabytes and stop the worker.
  • Make registration-gate detection linear in the length of the page. A page of many short paragraphs used to hold the worker for tens of seconds.

Fixed

  • Stop retrying a job that keeps stopping the worker. A job is put back in the queue at most 3 times after the worker dies without being asked to stop, then marked failed with its last error kept. A restart that follows a stop signal, such as a deploy, does not count. Migration 032_job_recovery_count adds a recovery_count column to jobs and changes no existing data.

0.63.3

Fixed

  • A title could still be announced twice when the repeated heading carried a markdown marker, such as "# Sunny", "## Sunny" or "Sunny". The check runs before leftover markers are removed, and it compared the heading with its marker attached, so short titles and bold titles did not match. The comparison now ignores heading and emphasis markers. Episodes narrated before this release keep the repeat until they are reprocessed.

0.63.2

Fixed

  • A book chapter still announced its title twice on 0.63.1. Cleanup can put the chapter number and the title on one line, such as "Chapter Six. A Quiet Start.", and 0.63.1 only compared whole lines. A heading line is now also checked part by part: the part that repeats the title is removed and the rest is kept, so that line becomes "Chapter Six." The same applies to a label and a headline on one line, such as "Technology - Headline". Episodes narrated before this release keep the repeat until they are reprocessed.
  • Three tests that upload the same book twice failed at random. The test EPUB builder stamped its zip entries with the current time, so two builds a few seconds apart were different files. It now uses a fixed timestamp. Uploads were not affected.

0.63.1

Fixed

  • Episodes could still announce their title twice. The intro reads the title, and the check that drops a repeated heading from the body only looked at the first block and only matched the whole title. It now checks the first five heading-like lines and stops at the first paragraph. It also matches one part of a title, such as the chapter in "Book: Chapter" or the headline in "Headline | Site", with or without a leading chapter number. A headline that shares a block with its subtitle is removed and the subtitle is kept. Episodes narrated before this release keep the repeat until they are reprocessed.

0.63.0

Added

  • Accept .epub uploads. Without the split option, the whole book is narrated in reading order as one episode.
  • Split an EPUB or PDF into one episode per chapter. POST /api/v1/upload takes two new form fields: split_chapters and chapters (all, default, or a list such as 1,3,5-9). Chapters come from the EPUB table of contents (EPUB 3 nav document, then EPUB 2 toc.ncx, then one chapter per spine document) or the PDF bookmark outline, top-level entries only. A chapter with a front or back matter title (copyright, contents, notes, index and similar) is unticked by default. A chapter shorter than MIN_EXTRACTION_CHARS cannot be selected.
  • Add the books API. GET /api/v1/books and GET /api/v1/books/{book_id} return a BookResponse with a status for each chapter (not_queued, queued, processing, done, failed, cancelled, removed). POST /api/v1/books/{book_id}/chapters queues chapters from all, default, remaining or a list of numbers, and reports skipped chapters that are already done or in flight. DELETE /api/v1/books/{book_id} returns 409 while any chapter is queued or processing.
  • Return chapters_detected: false from POST /api/v1/upload when split_chapters is set and the file has no chapters. The file is queued as one episode.
  • Add a "Split into chapters" checkbox to the Home upload area for EPUB and PDF files. Add a Books section with Choose chapters, Resume and Delete, and a chapter picker with Select all, Defaults and None.
  • Tag chapter episodes with itunes:season, itunes:episode, podcast:season (named after the book) and podcast:episode. The season is allocated per book and the episode number is the chapter's position in the full table of contents. The feed stays episodic.
  • Store each book's original and cover under {DATA_DIR}/books/{book_id}/, outside the media directory, so the orphan sweep leaves them alone. They stay until the book is deleted. Reprocess and requeue of a chapter episode read the source from there.
  • Reject EPUBs whose spine documents are encrypted (DRM) with a 400. Font-only encryption.xml entries are ignored. Archives are capped at MAX_UNCOMPRESSED_BYTES (200 MiB), MAX_ARCHIVE_ENTRIES (5000 entries) and MAX_EXTRACTED_CHARS (50 MiB of text).
  • Add chapter (the chapter number, or null) to GET /api/v1/jobs items and to upload webhook payloads. The Queue and Recents rows show "filename, chapter N" for a chapter job.
  • Add voice_id to the book response. The chapter picker shows "Voice: slot N (book's voice)" when Random is chosen and the book already has a voice.
  • Name a book after its filename, without the extension, when the EPUB or PDF has no title.
  • Pad a book cover to a square on the cover's average colour instead of cropping it. A small cover is used rather than falling back to the show image.
  • Pick up a deleted book's chapter episodes still in the feed when it is uploaded again. They show as done and move to the new season.
  • Add migration 031_books with the books and book_chapters tables and nullable season_number and episode_number columns on episodes. It is additive and changes no existing rows.

Changed

  • The queue claims jobs created in the same second in insertion order, so chapters queued together are narrated in chapter order.
  • A book narrates in one voice. The voice is stored the first time chapters are queued and reused on later calls unless voice is given, so a random default does not change voice between chapters.
  • Bump pydantic to 2.14.3 and ruff to 0.17.5 in all three packages, huggingface-hub to 2.2.0, and the uv build image and CI uv to 0.13.9.

Security

  • Accept three HIGH CVEs that Debian trixie has no fixed package for: CVE-2026-124709 (OpenSSL, all images), CVE-2026-109094 (X.Org server, render) and CVE-2026-109908 (libfontenc, render). The 0.62.1 app and render images carry the same findings. Each has a dated entry in the image's trivy ignorefile with a recheck date of 2026-12-22.

Don't miss a new Audicle release

NewReleases is sending notifications on new releases.