v0.8.1 — robust overlap finalization
Fixes & improvements
Overlap finalization + segment-timestamp dedup. Fixed-window chunk cuts land mid-word, which garbled weak models and could duplicate a phrase at the commit seam. The engine now commits finalized chunks by absolute segment timestamp and holds back the trailing stream_overlap seconds (cut mid-word) to be re-decoded with full context in the next chunk.
Changed
basestreaming is now robust and complete — recovers quiet tails a single batch decode dropped, and no seam duplication.- Small models (
tiny) go from garbled/hallucinated output to mostly coherent on fluent/pause-less audio. (Weak models can still hallucinate sporadically —base+ recommended for production.) - Repetition hallucinations are dropped (segment compression ratio > 2.4, Whisper's own failed-decode signal).
- New
LocalWhisperEngine.transcribe_segments; re-addedstream_overlap(default 3.0s).
See CHANGELOG.md.