The open-source software behind Weorclip
Weorclip is built on a handful of open-source components. Each one is used under its own licence; this page gathers where each comes from and what it is bound by.
Ships with the extension package
These files are distributed inside the program.
| Component | File / version | Licence |
|---|---|---|
| Fabric.js | 5.3.0 — fabric.min.js | MIT |
| Fabric.js EraserBrush | eraser_brush.mixin.js | MIT |
| Paper.js | paper-core.min.js | MIT |
| MP4Box.js (GPAC) | mp4box.all.min.js | BSD-3-Clause |
| MediaPipe Tasks Vision | mediapipe/vision_bundle.mjs | Apache-2.0 |
| MediaPipe BlazeFace model | mediapipe/blaze_face_short_range.tflite | Apache-2.0 |
| MediaPipe BlazeFace (full range) | mediapipe/blaze_face_full_range.tflite | Apache-2.0 |
| MediaPipe Pose Landmarker (lite) | mediapipe/pose_landmarker_lite.task | Apache-2.0 |
| SAM 2 — video predictor source code, 20 files, unmodified | host/sam2/ — github.com/facebookresearch/sam2 @ 2b90b9f5 | Apache-2.0 — © Meta Platforms, Inc. |
| LR-ASD active speaker detection model — AVA-ActiveSpeaker weights exported to ONNX, weights unchanged | host/lrasd_ava.onnx — github.com/Junhua-Liao/LR-ASD @ 1b6dcd2d | MIT — © 2025 Liao Junhua |
| AVA-ActiveSpeaker labels — training data of the LR-ASD weights, not shipped | research.google.com/ava | CC BY 4.0 — © Google LLC |
| YuNet face detection, uint8 input variant — weights unchanged | host/yunet_nhwc.onnx — from face_detection_yunet_2023mar.onnx (github.com/opencv/opencv_zoo) | MIT |
| Noto Emoji (40 PNG) | emoji/emoji_u*.png | SIL OFL 1.1 |
Downloaded at install — not bundled
The installer does not carry these tools; the user’s machine gets them straight from the publisher. Weorclip does not distribute them.
| Component | How it arrives | Licence |
|---|---|---|
| yt-dlp | winget install yt-dlp.yt-dlp, else github.com/yt-dlp/yt-dlp/releases/latest | The Unlicense (public domain) |
| FFmpeg / FFprobe | winget install yt-dlp.FFmpeg, else github.com/yt-dlp/FFmpeg-Builds → ffmpeg-master-latest-win64-gpl.zip | GPL-3.0 (the win64-gpl build) |
Why FFmpeg is listed separately. The FFmpeg build in use is GPL-licensed. Because Weorclip does not bundle FFmpeg — it calls the copy the user installed, as a separate process, over a command line — the GPL obligation does not reach Weorclip’s own source. This separation is kept on purpose.
Optional components — only if you ask
Models and tools that come down when features like subtitles, audio cleanup, speaker separation or translation are switched on.
| Component | Install script | Licence |
|---|---|---|
whisper.cpp (whisper-cli.exe, ggml*.dll) | host/motor_kur.ps1 → github.com/ggml-org/whisper.cpp/releases (newest release that carries a Windows build) | MIT |
Whisper GGML models (ggml-large-v3-turbo-q5_0.bin) | host/motor_kur.ps1 → huggingface.co/ggerganov/whisper.cpp | MIT (model weights: OpenAI, MIT) |
SDL2 (SDL2.dll, ships with whisper.cpp binaries) | host/motor_kur.ps1 (inside the whisper.cpp build) | zlib |
Silero VAD (ggml-silero-v5.1.2.bin) | host/motor_kur.ps1 → huggingface.co/ggml-org/whisper-vad | MIT |
llama.cpp (llama-server.exe, ggml*.dll; offline translation and smart captions) | host/motor_kur.ps1 → github.com/ggml-org/llama.cpp/releases (CPU build; CUDA 12.4 build on NVIDIA cards with ≥ 20 GB) | MIT |
TranslateGemma 4B (translategemma-4b-it.Q4_K_M.gguf) | host/motor_kur.ps1 → huggingface.co/mradermacher/translategemma-4b-it-GGUF | Gemma Terms of Use — see below |
TranslateGemma 12B (translategemma-12b-it.Q4_K_M.gguf, only on NVIDIA cards with ≥ 20 GB) | host/motor_kur.ps1 → huggingface.co/mradermacher/translategemma-12b-it-GGUF @ fdf84c9f, size + SHA-256 verified | Gemma Terms of Use — see below |
NVIDIA CUDA runtime (cudart64_12.dll, cublas64_12.dll, cublasLt64_12.dll, with the llama.cpp CUDA build) | same llama.cpp release → cudart-llama-bin-win-cuda-12.4-x64.zip | NVIDIA CUDA Toolkit EULA (redistributable components) |
| DeepFilterNet (noise removal) | host/motor_kur.ps1 → github.com/Rikorose/DeepFilterNet/releases (Windows .exe) | MIT / Apache-2.0 (dual) |
WeSpeaker model (wespeaker_en_voxceleb_resnet34_LM.onnx) | host/konusmaci_kur.ps1 → github.com/k2-fsa/sherpa-onnx | Apache-2.0 |
YuNet face detection (face_detection_yunet_2023mar.onnx) | host/gorme_kur.ps1 → github.com/opencv/opencv_zoo | MIT |
YOLOX-tiny object detection (yolox_tiny.onnx; smart framing: objects, screens, cartoon characters) | host/gorme_kur.ps1 → github.com/Megvii-BaseDetection/YOLOX/releases (0.1.1rc0), size + SHA-256 verified | Apache-2.0 — © Megvii Inc. |
ONNX Runtime DirectML (onnxruntime-directml; runs the smart framing models on the graphics card, or on the processor) | host/gorme_kur.ps1 → pip install onnxruntime-directml into <install>\vision\venv (processor build onnxruntime only as a fallback) | MIT — © Microsoft Corporation. The wheel also contains DirectML (DirectML.dll, Microsoft DirectML licence terms) — installed with pip on the user machine, not redistributed by Weorclip |
Qwen3.5-4B (Qwen3.5-4B-Q4_K_M.gguf, smart captions: emphasis word, emoji, hook title) | host/motor_kur.ps1 (component akilli) → huggingface.co/unsloth/Qwen3.5-4B-GGUF (base: Qwen/Qwen3.5-4B), SHA-256 verified | Apache-2.0 — © Alibaba Cloud (Qwen team) |
MODNet portrait matting (modnet.onnx, hook title behind the speaker) | host/motor_kur.ps1 (component akilli) → huggingface.co/Xenova/modnet (ONNX export of github.com/ZHKKKe/MODNet), SHA-256 verified | Apache-2.0 — © ZHKKKe |
| Demucs (audio separation) | host/motor_kur.ps1 (component sesayir) → pip install demucs numpy in its own venv | MIT |
| Python (only if audio separation is chosen) | host/motor_kur.ps1 → winget install Python.Python.3.12 | PSF License |
TranslateGemma is not open source. Gemma-family weights are distributed under Google’s Gemma Terms of Use and a Prohibited Use Policy — not MIT or Apache. The model is downloaded only if you ask and is never redistributed by Weorclip, but the terms still bind you.
Voice engine — read-aloud and dubbing
This engine is optional too: it is installed only when you select it, or the first time you use read-aloud or dubbing. Every file is pinned by revision, size and SHA-256, and Weorclip redistributes none of it.
| Component | How it arrives | Licence |
|---|---|---|
VoxCPM2 model (model.safetensors, audiovae.pth, tokenizer) and the voxcpm 2.0.3 package | host/motor_kur.ps1 → huggingface.co/openbmb/VoxCPM2 @ 32279eff · pip install voxcpm==2.0.3 --no-deps | Apache-2.0 — © OpenBMB; the model repository has no LICENSE file, the licence is declared by the upstream project |
Mel-Band RoFormer vocal separation weights — MelBandRoformer.ckpt | host/motor_kur.ps1 → huggingface.co/KimberleyJSN/melbandroformer @ ac9b0614 | MIT — © Kimberley Jensen |
| python-audio-separator 0.47.0 | host/motor_kur.ps1 → pip install audio-separator==0.47.0 --no-deps | MIT — © Andrew Beveridge |
| PyTorch 2.8.0 — CUDA 12.8 or CPU build | host/motor_kur.ps1 → download.pytorch.org | BSD-3-Clause |
| torchaudio 2.8.0 | host/motor_kur.ps1 → download.pytorch.org | BSD-2-Clause |
| NVIDIA CUDA runtime libraries — only inside the CUDA build of PyTorch | host/motor_kur.ps1 → same wheel as PyTorch | NVIDIA CUDA Toolkit EULA — redistributable components |
| Transformers, safetensors, huggingface_hub, tokenizers | host/motor_kur.ps1 → PyPI, pinned | Apache-2.0 |
| librosa | host/motor_kur.ps1 → PyPI, pinned | ISC |
| NumPy, SciPy, scikit-learn, soundfile | host/motor_kur.ps1 → PyPI, pinned | BSD-3-Clause |
| Numba, llvmlite | host/motor_kur.ps1 → PyPI, pinned | BSD-2-Clause |
| ONNX Runtime, pydub, einops | host/motor_kur.ps1 → PyPI, pinned | MIT |
| python-soxr — resampling, pulled in by librosa | host/motor_kur.ps1 → PyPI, pinned | LGPL-2.1-or-later — installed as a separate, replaceable package in the user’s own environment; not linked into Weorclip |
| certifi — CA certificate bundle | host/motor_kur.ps1 → PyPI, pinned | MPL-2.0 |
| Python 3.12.10 — only if no suitable Python exists | host/motor_kur.ps1 → winget --scope user, else python.org installer (SHA-256 + Authenticode check) | PSF License |
Dubbed videos are labelled. A dubbed video carries a note in its metadata saying its audio was dubbed with AI, plus the language tag of its audio track. The VoxCPM2 model card forbids impersonation and fraud, so Weorclip offers only designed voices and never clones anyone else’s voice.
Object tracking engine — Super Edit
The SAM 2 source code ships with the program (Apache-2.0, unmodified). The model and a few small helper packages download at pinned versions the first time you start object tracking, and Weorclip does not redistribute them. The engine shares the voice engine’s Python environment, so the AI runtime is never downloaded twice.
| Component | How it arrives | Licence |
|---|---|---|
SAM 2.1 Hiera-Tiny model — sam2.1_hiera_tiny.pt | host/motor_kur.ps1 → dl.fbaipublicfiles.com, size + SHA-256 verified | Apache-2.0 — © Meta Platforms, Inc. |
| Hydra (hydra-core 1.3.7), iopath 0.1.10 | host/motor_kur.ps1 → PyPI, pinned | MIT |
| OmegaConf 2.3.1, ANTLR 4 Python runtime 4.9.3, portalocker 4.4.0 | host/motor_kur.ps1 → PyPI, pinned | BSD-3-Clause |
| Pillow 12.3.0 | host/motor_kur.ps1 → PyPI, pinned | MIT-CMU (HPND) |
| OpenCV — opencv-python-headless 5.0.0.93 | host/motor_kur.ps1 → PyPI, pinned | Apache-2.0; the wheel contains FFmpeg as a separate DLL (LGPL-2.1), installed in the user’s own environment |
| PyTorch, NumPy | shared with the voice engine above — not downloaded twice | BSD-3-Clause |
Third-party services
Gemini, OpenAI and Anthropic come into play only when you enter your own key and run the feature. Weorclip operates no server of its own.
Licence texts
The full text of each licence is in the licenses/ folder of the install directory and in the THIRD-PARTY-NOTICES.md file next to the program.
This page was prepared from each component’s licence headers and upstream repository. It is not legal advice.
← Back to the home page