Changelog

[3.2.0] - 2026-06-20

Web Worker offloading — NLP computation (TF-IDF, cosine similarity, TextRank, POI detection) now runs off the main thread
Multi-language stopword support with auto-detection for 10 languages (es, fr, de, pt, ja, ko, zh, hi, ar, ru)
TextTiling depth-score boundary detection replaces raw cosine similarity thresholds for higher-confidence chapter boundaries
Adaptive NLP window sizing — 120s groups for videos >45min, 60s for shorter
Intra-segment pause detection from word-level timing gaps
IndexedDB caching with LRU eviction (replaces localStorage, auto-migrates existing cache)
Settings export/import via JSON files
Chapter quality confidence score (high/medium/low) shown in Chapters tab
Transcript download in SRT and VTT formats
Chapter export as JSON with structured data (timestamps, titles, POIs, confidence)
Undo button on auto-skip toasts to seek back to pre-skip position
Filler word regex compilation memoization
Multi-language filler word catalogs (es, fr, de, pt, ja) — auto-activated when non-English transcript detected
SponsorBlock integration — fetches segment data and snaps chapter boundaries to sponsor/intro/outro edges

Chapter tracking and auto-skip merged into single RAF playback loop (halved per-frame CPU cost)

Implemented missing _gmGet, _gmPostJson, and _buildSapisidAuth methods that caused transcript fallback methods 4-6 to throw TypeError (dead code since v3.0.0)
Added abort controller to transcript fetch pipeline — rapid SPA navigation no longer causes stale results from a previous video to overwrite current data
Added cache schema validation — corrupt or version-mismatched cached data is silently discarded and regenerated instead of crashing