SubbyAI 1.0.2

SubbyAI 1.0.2 — Windows

Verified source revision: 14f7616aa90272936c132ab23d605459b70789d6.

The Overlay button now acknowledges a click immediately and the subtitle bar paints a visible ready, starting, downloading, waiting or error state even before speech arrives. Capture startup no longer determines whether an enabled overlay gives feedback. Slow settings writes and opt-in history initialization/finalization run on ordered workers; late lifecycle events cannot reopen the application during Quit. Native window recovery preserves focus and uses the same overlay window.

The accuracy investigation found pipeline faults including inconsistent resampling, overly aggressive speech splitting, lost boundary context, independent chunk recognition, overlap duplication, fragment translation and unavailable confidence treated as zero. Continuous preprocessing, padded pause-aware segments, bounded recognition/context history, provisional revisions, timestamp/text reconciliation and finalized sentence translation preserve more of the models' capabilities. Late translations cannot cross history ownership boundaries, and Quit drains the final buffered words before storage closes.

Validation includes 872 passing tests, Ruff, compilation/version/lock checks, current dependency advisories, source/history secret scans, and byte-for-byte comparisons of original, processed and exact ASR chunk audio. The fixed 14-fixture production-small regression corpus had no literal WER/CER regression against the old live pipeline; quiet and antiphase controls improved from 100% WER to 0%, and human English from 13.64% to 0%. Rare names, competing speakers and multilingual translation still have model/evaluation limits; this is a regression corpus, not a general accuracy guarantee.

Installed production verification: The installed 1.0.2 production build matches all 436 packaged files and the verified commit. Its actual native Windows check passed 20 full ON/OFF cycles and 47 timed clicks (maximum handler time 7.6 ms), including blocked initialization, slow settings/history, failure/retry, window recovery and button/bar agreement. CUDA Whisper, CPU Parakeet and local English-to-French translation passed on prerecorded input. Normal installed-binary launch and graceful shutdown passed in an isolated profile; clean first-run installation and high-DPI returning-user startup also passed. Earlier six manual ordinary-profile cycles verified the pink indicator and bar; Overlay/runtime code is unchanged since that candidate apart from the version constant. No speakers or physical recording were used.

Windows 11 25H2 was the available test environment, accepted by the project owner for this issue. After the owner's request, checks used prerecorded files and simulated capture without speaker playback or physical recording. Earlier hosted Actions were unavailable. Subsequent hosted failures blocked publication and were investigated; The complete CI and Release workflows subsequently passed on Windows and both macOS platforms, including draft generation. Windows release gates were also run locally. No unverified macOS binaries are included.

Windows binaries are unsigned; see the installation guide.

Download the setup installer for a per-user installation, or the portable archive. Settings, history and downloaded models remain outside the application directory. SHA256SUMS-windows.txt identifies the three distribution assets; the corresponding Qt/PySide source archive and license notices accompany this release.

See the root causes, evidence, before/after metrics and remaining limits and the complete accuracy investigation.

The 1.0.1 candidate was withheld after hosted tests exposed queued-signal and animation timing assumptions. Version 1.0.2 corrects those tests and strengthens platform recovery coverage; no tests were disabled and the earlier tag was not moved. The original detailed evidence remains applicable to the unchanged caption pipeline.