Skip to content

PolyVox Studio v2.8

Choose a tag to compare

View all tags
@dmarsh400 dmarsh400 released this 18 May 14:42
· 7 commits to main since this release
v2.8
88a06c7

PolyVox Studio v2.8

Release date: 2026-05-18

Highlights

  • IMPROVED: Character detection accuracy — significantly reduced misattribution in first-person and dialogue-heavy chapters.
  • IMPROVED: Main-character inference now uses address-based scoring to correctly identify the POV protagonist across chapters.
  • FIXED: Brand/object names (e.g. "Smith & Wesson") no longer leak into speaker attribution as fake characters.
  • IMPROVED: EPUB chapter detection now correctly splits chapters within parts instead of treating parts as the top-level unit.

Detailed Changes

Character Detection (app/core/character_detection.py)

New: Address-based scoring for main-character inference

The _infer_main_character() function now tracks how often each character is addressed by name inside dialogue lines, in addition to the existing narration-mention and quote-count signals. Being directly addressed in dialogue ("You're a lucky man, Kovacs.") is a very strong signal that a character is the protagonist. This new addressed score is weighted at 6× and an additional 2× for first-person books, which reliably lifts the real POV character above secondary characters in the scoring.

Before this change, the prologue of Altered Carbon scored Sarah above Takeshi Kovacs because Sarah had more raw narration mentions. After this change, Chapter One scores Takeshi Kovacs at 118.5 vs Warden Sullivan at 50.0.

Improved: Book-level main-character aggregation in GUI

The _resolve_main_character_override() method in the GUI now sums the actual candidate scores from each chapter instead of summing the per-chapter confidence values. This means a character who narrowly wins one chapter but is clearly dominant in another now wins at the book level.

Tuned: Scoring weights

Signal Old weight New weight
Narration mentions 5.0 (+ 2.0 fp) 4.0 (+ 1.5 fp)
Quote count 2.0 1.5
Address mentions — 6.0 (+ 2.0 fp)
Mention count (unknown ratio) 0.5× 0.25×

Character Attribution UI (app/ui/characters_tab.py)

Fixed: "Smith & Wesson" → "Wesson" false speaker

The explicit-character-from-context matcher now skips any name match where an ampersand (&) appears immediately to the left of the matched name in the surrounding text. This suppresses brand names like Smith & Wesson and Heckler & Koch from being promoted as real character speakers.

Improved: Book-level main-character override aggregation

_resolve_main_character_override() now sums raw candidate scores (from all chapter-level inferences) instead of chapter-level confidence floats, giving a more representative book-wide ranking when presenting the override dropdown to the user.

EPUB Chapter Detection (app/core/chapter_chunker.py)

Fixed: Chapters nested inside Parts are now detected correctly

Previously, EPUB files that organised content as PART I → Chapter 1, Chapter 2 … were splitting only on the Part markers, producing very large, undivided segments. The chapter chunker now walks the spine in order and detects chapter-level headings within parts, falling back to size-based chunking only when no chapter markers are found. Word-number labels ("Chapter One", "Chapter Two", …) are also recognised.

Migration / Cache Notes

  • The BookNLP cache key is unchanged (v2). Existing caches remain valid.
  • No changes to the voices or synthesis pipeline.
  • No database or settings file migrations required.
Assets 2
Loading