Skip to content

v1.2 — Script Review Pass & Bug Fixes

Choose a tag to compare

View all tags
@Finrandojin Finrandojin released this 09 Feb 01:17
· 142 commits to main since this release
v1.2
07a58a8

Alexandria v1.2 — Script Review Pass & Bug Fixes

A new LLM-powered script review feature that catches common annotation errors, plus critical fixes for CPU users and Windows audio export.

LLM Script Review (New)

  • One-click review pass — "Review Script" button in the Script tab runs a second LLM pass over the generated script to detect and fix common annotation errors
  • 5 automated fix categories:
    1. Attribution tags stripped from dialogue ("said he", "she replied")
    2. Narration mixed into character entries split out as NARRATOR
    3. Dialogue embedded in narrator entries extracted as the correct speaker
    4. Short consecutive narrator entries merged for consistent TTS delivery
    5. Invalid instruct fields (physical actions) replaced with voice directions
  • Text-loss safety net — Programmatic validation rejects any batch where significant text was dropped, falling back to the original entries
  • Customizable prompts — Review prompts stored in review_prompts.txt (same separator format as generation prompts), hot-reloaded on every request
  • Batch processing with context — Processes script in batches of ~25 entries, passing tail context between batches for accurate speaker attribution at boundaries

Bug Fixes

  • Fix CPU model loading crash — device_map="cpu" triggered HuggingFace accelerate's meta tensor dispatch, causing "Cannot copy out of meta tensor; no data!" errors. Now omits device_map on CPU. (Fixes #4)
  • Fix parallel generation race condition — Multiple TTS worker threads were overwriting each other's status updates to chunks.json, causing the UI to show random Error/Pending states despite all chunks succeeding. New atomic _update_chunk_fields() method ensures thread-safe read-modify-write. (Fixes #3)
  • Fix silent MP3 encoding failure — Conda's Windows ffmpeg often lacks libmp3lame, causing pydub.export() to silently produce a 428-byte header-only MP3 file. Now validates exported file size and automatically falls back to WAV when MP3 encoding fails. (Fixes #3)

Prompt Architecture

  • Single source of truth — All default prompts consolidated into default_prompts.txt (generation) and review_prompts.txt (review) at project root
  • Hot-reloadable — Edit prompt files and changes take effect immediately, no app restart needed
  • Improved generation prompts — Refined narrator/character split rules and instruct directions based on iterative testing
Assets 2
Loading