0.1.0 pre.4

Pre-release of Voxprint AI Movie Dubber. This is an early version: bugs are still being fixed, but it already shows how the app works.

Small online installer (Windows 10/11 x64): Python, PyTorch and the other components are downloaded during setup. An NVIDIA GPU is needed for useful speed.

What's new - Fix: the installer now detects CUDA on new NVIDIA drivers (6xx, "CUDA UMD Version" in nvidia-smi) and installs the GPU build of PyTorch; v0.1.0-pre.3 installed the CPU build there. - Lines follow sentences: recognised speech is cut at sentence ends, pauses over 0.6 s and speaker changes instead of every 4-8 s; punctuation is kept for the whole film. - Complete translations: every sentence is translated and checked; short exclamations such as "Learning sucks!" or "Destroy!" are no longer dropped. - Time fitting keeps the meaning: a line is shifted into a pause or sped up (up to 1.15x) before it is shortened, and a shortened line keeps at least 70 % of the text. - Switching between MKV and MP4 only rebuilds the output file. - Speech recognition runs on the GPU: PyTorch 2.11.0 (CUDA 12.8) is pinned like in Voxprint AI Audiobook Builder. - Works alongside Voxprint AI Audiobook Builder: shared runtime (when the pinned versions match), shared settings (suite.json), shared Start menu folder "Voxprint". Uninstalling never touches the Audiobook Builder. - Speech the recogniser skipped is recovered with a second pass; hallucinated burst lines and stray marks are removed. - License: Apache-2.0 (also shown by the installer).

Verify the download with VoxprintMovieDubber-Setup.exe.sha256.