Local speech to text Qwen3-ASR w/ OpenVINO (no API key)
v1.0.1Local offline ASR on Windows — no cloud, no API cost, full privacy. Qwen3-ASR 0.6B + Intel OpenVINO, GPU-accelerated inference. NETWORK: required for first-t...
⭐ 0· 91·0 current·0 all-time
by@juan-oy
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name/description (local Qwen3-ASR with OpenVINO) matches the code and instructions: setup.py creates a venv and installs packages, download_model.py pulls the Qwen3-ASR ModelScope snapshot, and runtime scripts perform extraction and offline inference. Required operations (git clone, pip install, modelscope download) are expected for this purpose.
Instruction Scope
SKILL.md instructs the agent to run setup.py and download_model.py for first-time setup and to use the provided transcribe/acoustic pipeline for inference. The scripts scan drives A:..Z: to find state.json and write files under a per-user folder (e.g., {USERNAME}_openvino). The README mentions proxy detection via HTTP_PROXY/HTTPS_PROXY and WinHTTP (netsh), but the shipped scripts implement Modelscope download and venv handling (they do check env vars but do not run netsh proxy probing); this is a minor doc/implementation mismatch. All runtime actions (cloning, pip installing, downloading ~2 GB) are within scope for installing a local ASR.
Install Mechanism
There is no registry install spec; the skill is instruction+scripts. setup.py performs venv creation, pip installs packages (openvino, modelscope, etc.), and git clone from GitHub; download_model.py uses modelscope.snapshot_download to fetch the model. These are standard mechanisms for model+tool installation. Network usage for initial setup and model download is explicit and expected. No unknown or obfuscated remote download URLs are used (ModelScope and GitHub are standard).
Credentials
The skill requests no credentials or secrets. It reads USERNAME and looks for system Python/VENV paths and will write state.json into a per-user folder. Network access is required only for setup and model download (as documented). No unrelated environment variables, cloud keys, or tokens are requested or embedded.
Persistence & Privilege
The skill writes persistent files to disk (creates {USERNAME}_openvino/ asr/venv/, state.json, outputs, and model files) and installs packages into a venv. This is expected for a local ASR. always:true is not set. Consider that these filesystem changes are persistent and potentially large (~2+ GB); the skill does not modify other skills or global agent configs beyond writing its own state in the chosen user-root folder.
Scan Findings in Context
[pre-scan-none] expected: No regex-based pre-scan findings were flagged. The observed behaviors (git, pip, modelscope download, reading USERNAME, scanning drives) are present in source and are expected for installing and locating a local model and venv.
Assessment
This skill appears to do what it claims: it will create a local Python venv, pip-install packages (including OpenVINO and modelscope), clone Qwen3-ASR from GitHub, and download ~2 GB of model files to a per-user folder (e.g., C:\<drive>\<USERNAME>_openvino\asr). Before installing or letting the agent run auto-bootstrap: 1) Be aware it will perform network activity and consume disk (≈2+ GB); ensure you have space and bandwidth. 2) Run setup.py and download_model.py manually in a terminal the first time if you prefer to review output and errors rather than letting the agent run them autonomously (auto-bootstrap will run these scripts automatically). 3) The code scans drives A:..Z: to find/write state.json and model files — expect files under a root-level per-user folder. 4) If you need privacy, the skill is designed for offline inference after setup; however the initial setup contacts ModelScope and GitHub. 5) If you don’t trust the source, inspect the GitHub repo QwenLM/Qwen3-ASR and the models on ModelScope, or run setup/download in an isolated VM. 6) If you want to avoid persistent changes, do not enable auto-bootstrap and run the install steps manually in a controlled environment.Like a lobster shell, security has layers — review code before you run it.
latestvk97cx0a6jk7yvkb5qkraca0nz583p4kh
License
MIT-0
Free to use, modify, and redistribute. No attribution required.