Skip to content

Changelog

All notable changes to Kenkui are recorded here. The format follows Keep a Changelog, and versions follow Semantic Versioning.

Unreleased

10.2.1 - 2026-11-02

Fixed

  • Preserve leading-dot decimal quantities during pronunciation: .085 is read as zero point zero eight five, and currency, percentages, units, and negatives retain their fractional value. Declined decimal forms cannot fall through to a whole-integer or year rewrite. The spoken-form recipe is now v2 so cached audio from the previous number rules is not reused.
  • Pocket-TTS synthesis now enforces its 50-token text budget on both the actual inference input and the prepared prompt. Long clauses fall back to complete word boundaries after sentence/clause boundaries, preserving text order. Audio cached with the previous unbounded splitter is regenerated.
  • Pin the optional PDF layout backend to RapidOCR 3.9.2 so fresh installs can resolve English OCR models with Docling 2.128.0. RapidOCR 3.10 removed the language-table API that this Docling release requires.

  • Configure full-page OCR through Docling's typed OcrMode.FULL_PAGE API instead of its legacy force_full_page_ocr compatibility keyword.

  • Normalize word-internal curly apostrophes immediately before Pocket inference so contractions use the tokenizer's trained apostrophe token. Canonical text stays unchanged, and earlier cached audio is regenerated.

  • Expose PdfLayoutOptions(ocr_use_cls=False) for verified upright scans whose lines the OCR orientation classifier incorrectly rotates. The default remains enabled for mixed-orientation documents.
  • Preserve PDF worker errors when an exiting process group rejects a shutdown signal, while still requiring surviving OCR processes to terminate.
  • Require writable segment storage when keep_audio_cache=True or previews request retention. Unavailable storage raises audio_cache_unavailable instead of silently discarding reusable audio. Operators can set an absolute private cache directory with KENKUI_AUDIO_CACHE_DIR.

10.2.0 - 2026-09-29

Changed

  • New synthesis pipelines announce authored chapter titles by default. Use chapter_titles(enabled=False) to opt out, or configure title pauses and spoken-only overrides. Matching opening headings are reused; source text and character offsets remain intact.

  • EPUB table-of-contents targets now define logical chapters, joining internal file splits and preserving unlisted content in reading order. Fragment targets can delimit multiple chapters in one file. Without usable navigation, use opening headings, meaningful titles, or sequential Untitled section N labels. Internal epigraph credits no longer name whole sections. Re-inspect sources and refresh chapter selections after upgrading: regrouped chapters have new IDs.

Added

  • Initial native PDF source support via pdf() or book(), with explicit pdf_processing(mode="native"), immutable prepare() checkpoints and pdf_report() evidence. Install the optional pdf extra. Preparation runs automatically before PDF resolution/rendering; inspection requires preparation.
  • Independently callable header/footer removal and paragraph reconstruction in kenkui.pdf_processing, versioned recipes, edit logs, source-change guards and preparation-aware plan/attribution identities. Empty recipes disable cleanup. Persistent caches are not yet available.
  • Extend the native PDF recipe with source-attested line-break word repair and conservative same-page footnote/reference removal. Unknown spellings remain unchanged and are reported; word repair preserves all letters.
  • Recognize corroborated top-of-page folios and numbered running titles, including facing-page layouts. Years and numbers inside prose are retained.
  • Add optional pdf-layout extraction through an isolated, bounded CPU worker, with provisioned local Docling/RapidOCR assets, selective OCR, native-text completeness checks, progress stages and strict result validation.
  • Add independently callable numbered note-section and visual-material removal, verified footnotes within layout paragraphs, and repeated OCR furniture removal.
  • Validate PDF recipe prerequisites before extraction, reject invalid resource limit types, and honor cancellation before cleanup callbacks run.

Fixed

  • Incomplete quote-attribution responses are retried and rejected before caching; old response and attribution caches are revalidated for full quote coverage.
  • Cached model responses now honor cancellation before reading and after validation.
  • Failed model calls report bounded diagnostics without exposing book text.

10.1.1 - 2026-09-20

Fixed

  • A scene ornament is no longer read aloud. A book that divides its scenes with * * *, #, or a similar glyph run on a line of its own was handing that run to the engine, which said it however it saw fit. Such a line is now recognized from the canonical text and speaks as nothing, and the block after one opens a scene, so pauses(scene_ms=...) covers these books too without the publisher having labelled anything. Suppression does not depend on pronounce(): a pipeline that never asked for a spoken form is exactly the one affected.

Canonical text is untouched, so billing, chapter identity, offsets and sidecar anchors are unchanged. The segments that held an ornament do change text, so those -- and only those -- re-synthesize once. A chapter containing nothing but an ornament is a separator page rather than a scene break and is left exactly as it was.

10.1.0 - 2026-09-20

Added

  • pauses(scene_ms=...) gives mid-chapter scene breaks their own duration. A scene break is detected at parse time from an <hr/> or a block the publisher labelled as one (class/epub:type, including an ornament hidden from assistive technology); normalization erases the distinction later, so it cannot be recovered afterwards. The tier is off unless set, so no existing render changes. ChapterInspection.scene_ranges reports the detected blocks and ScriptRow.is_scene_start marks the row that opens a scene, while the silence it implies falls on the row before it.

10.0.0 - 2026-09-20

Kenkui 10 is a ground-up rewrite. It shares no code with the 2.x series, and none of the 2.x API carries over. The version is 10 in binary: the second generation.

Added

  • An immutable, typed Pipeline API: book(), fluent intent methods, validate(), inspect(), resolve(), and write(), plus magic_run() for one-call rendering.
  • Local speech synthesis with Pocket-TTS 2.1.0, in spawned worker processes, with a private per-book audio cache and atomic M4B publication through FFmpeg.
  • Explicit voice provisioning with load_voice(), unload_voice(), add_voice(), remove_voice(), and list_voices(), over a built-in catalog of 26 upstream voices and 95 precompiled kenkui-voices.
  • Multi-voice casting: character discovery with spaCy or a LiteLLM model, an optional identity pass that merges aliases, LiteLLM quote attribution (including unnamed speakers), and a deterministic casting solver with gendered and random methods.
  • Character review with resolve(until="characters") and with_characters().
  • Series continuity: .series() keeps characters' voices across volumes.
  • Speech shaping: pronounce() with lexicons, number-reading tiers, and vocal gesture collapsing; pauses(); and cover art control.
  • The dial-in loop: script(), attribute(), silence(), scoped pronounce(), select(), preview(), and sidecar corrections with annotations() and write_annotations().
  • Progress events, cooperative cancellation, and stable ErrorCode values on every failure.
  • Runnable examples in examples/.

Fixed

  • Quote attribution now returns gender directly in a separate per-speaker table. Valid, consistent evidence survives storage and offline roster refreshes; conflicting evidence is flagged and reviewed genders retain precedence. A controlled four-book passage evaluation found no attribution regression when this field was added to the role-aware prompt; see the evaluation report for sample limits and provider-routing anomalies.
  • Character attribution now requests distinct, gender-qualified identities for unnamed people sharing a role, using pronouns and actions in context. Casting recognizes those qualifiers; sparse unopposed dialogue tags fill unknown genders, inverted tags and common adverbs are recognized, and ambiguous tag evidence is logged. The prompt version advances to avoid reusing old merges.
  • Chapter names now prefer EPUB table-of-contents labels over epigraph headings and internal document titles. Calibre split continuations inherit the label with a part number; filename-only titles use the numbered fallback.