Use macOS or Linux for version 0.5.1. Windows is planned for a later release and is untested. Linux with NVIDIA is assumed, not verified. No Linux hardware receipt exists. Local macOS tests cannot prove Linux CUDA loading or NVENC output.
On macOS, install ffmpeg through Homebrew:
brew install ffmpegOn Ubuntu or Debian:
sudo apt-get update
sudo apt-get install -y ffmpegFor other Linux distributions, use your package manager. Confirm both tools:
ffmpeg -version
ffprobe -version
declip doctor --jsonYour build must provide libx264, libx265, and an AAC encoder for software video renders.
Doctor exits 1 when a tool or transcription backend is unavailable.
A core-only editing install can still review, render, and export existing plans.
Install the mac extra for MLX transcription on macOS 14 or later.
The locked MLX and Torch wheels require that minimum version.
The Mac-extra audit uses the same macOS 14 deployment target.
You get software decode unless a source-specific hardware decode probe passes.
A listed hardware encoder must also pass an encode probe before selection.
The local render test verified HEVC Main 10 through VideoToolbox.
Your output also passes a codec and bit-depth check after rendering.
Install the cuda extra.
You need a compatible NVIDIA driver, CUDA 12 libraries, and cuDNN 9 libraries.
The extra supplies cuBLAS and cuDNN wheels.
The loader opens libcublas.so.12 and libcudnn*.so.9 from their installed package directories before importing CTranslate2.
This preload path remains unverified on Linux hardware.
Check nvidia-smi and declip doctor --json before your first run.
An explicit --device cuda fails if CUDA is unavailable.
Use --device cpu with a CPU backend when you choose to run on CPU.
Windows CUDA DLL loading is outside version 0.5.1. Do not interpret wheel availability as Windows support.
On the remote machine:
declip review clip.mp4 --no-browser --port 8765On your local machine:
ssh -L 8765:127.0.0.1:8765 your-hostOpen the printed http://127.0.0.1:8765/?token=... URL locally.
Keep the same port at both ends so the Host check passes.
Keep the token private and leave SSH connected until you finish review.
The server binds only to loopback.
You cannot select a public bind address.
Use none to preserve PCM without denoising.
Use afftdn for the built-in ffmpeg denoiser.
With auto, you get deepfilter if available, otherwise afftdn.
An explicit unavailable deepfilter choice fails.
You must install the standalone deep-filter binary yourself.
The app does not download it.
Put it on PATH as deep-filter, or set deepfilter_path in your config.json.
P6 inspected the DeepFilterNet v0.5.6 assets on 2026.10.06.
| Platform | Standalone release asset | Evidence |
|---|---|---|
| macOS arm64 | deep-filter-0.5.6-aarch64-apple-darwin |
Asset and local stereo probe verified |
| macOS x86_64 | deep-filter-0.5.6-x86_64-apple-darwin |
Asset verified; execution unverified |
| Linux x86_64 | deep-filter-0.5.6-x86_64-unknown-linux-musl |
Asset verified; execution assumed |
| Linux aarch64 | deep-filter-0.5.6-aarch64-unknown-linux-gnu |
Asset verified; execution assumed |
| Linux armv7 | deep-filter-0.5.6-armv7-unknown-linux-gnueabihf |
Asset verified; execution assumed |
| Windows | Release asset exists | Unavailable in 0.5.1; planned and untested |
| Other combinations | No selected asset | Unavailable; use afftdn |
Strength maps from zero to one onto --atten-lim-db from zero to one hundred dB.
The enhancer restores your source sample rate, channels, and sample count after processing.
The upstream neural binary uses a 16-bit intermediate.
A final 24-bit WAV cannot recover precision lost inside that binary.
Hardware certification requires separate platform receipts.
Use the README evidence matrix for the current stated limits.
GitHub Actions runs the test suite on ubuntu-latest and macos-latest for each push. These runners have no GPU.
| Cell | State | Evidence or gap |
|---|---|---|
macOS arm64: locked mac extra and MLX on Metal |
Verified | Metal receipt, passed. Cached large-v3-turbo weights, real backend, 18/18 filler labels, zero false hits, real-time factor 0.060. |
| macOS arm64: VideoToolbox H.264 and HEVC Main 10 | Verified | Metal receipt: H.264 yuv420p; HEVC yuv420p10le, profile Main 10. |
| macOS arm64: VideoToolbox A/V sync | Verified | Metal receipt: 250 keeps, 250 flashes, 250 beeps. Maximum offset 0.498 frames, including batch borders 40/41 and 80/81. Duration error 0.014 frames. |
CPU on macOS: locked cpu extra and faster int8 |
Verified | CPU receipt, passed. Cached large-v3-turbo weights, 16/18 filler labels, one false hit, real-time factor 0.228. CUDA device count 0. |
| CPU on macOS: software H.264 and HEVC Main 10 | Verified | CPU receipt: H.264 yuv420p; HEVC yuv420p10le, profile Main 10. Software encoding was forced. |
| CPU on macOS: software A/V sync | Verified | CPU receipt: 250 keeps, 250 flashes, 250 beeps. Maximum offset 0.498 frames, including batch borders 40/41 and 80/81. Duration error below 0.001 frames. |
| Linux with NVIDIA: CUDA loading, transcription, NVENC, and sync | Assumed; receipt waived for 0.5.0 | Verification hardware is unreachable. Ruling 11 (2026.10.06) waives the Linux receipt. No remote verification attempted. |
| CPU on Linux | Assumed | The CPU receipt measures this Mac only; no Linux execution receipt exists. |
| Windows | Deferred; planned and untested | Ruling 10 (2026.10.06) moves Windows to a later release. No Windows verification attempted. |
| Prompt persistence after thirty seconds | Unproven | These smoke receipts measure prompted synthetic recall. They do not compare a no-prompt baseline or establish real-speech persistence. |
| Hosted CI | Verified | GitHub Actions passed all jobs on ubuntu-latest and macos-latest (run 37505366506). These runners have no GPU. |