v0.8.4

Changed

• Audio language detection samples multiple positions. A single 30s window at 33% into a track could land on music/action/silence and yield a low-confidence guess that failed the gate (observed: a 5.1 track returning pt@0.40, left und even though it has speech elsewhere). Detection now tries several spread-out windows (33%/55%/15%/72%) and takes the most confident, short-circuiting as soon as one clears the threshold — so clear tracks still cost a single sample while weak-first-window tracks get detected from a better window. Applies to on-demand and pre-conversion audio detection.

Docker images

Pinned to this release:

docker pull ghcr.io/i-ial9000/shrinkerr:0.8.4         # portable (amd64 + arm64, CPU)
docker pull ghcr.io/i-ial9000/shrinkerr:0.8.4-nvenc   # x86_64 + NVIDIA GPU (NVENC)
docker pull ghcr.io/i-ial9000/shrinkerr:0.8.4-edge    # as above but ffmpeg master
docker pull ghcr.io/i-ial9000/shrinkerr:0.8.4-edge-nvenc

Floating tags (:latest, :edge, :nvenc, :edge-nvenc) track the most recent release of each lineage.