Qwen ASR

Local speech-to-text using Qwen3-ASR (CPU-only, no API key, no cloud). Use when: (1) a voice message or audio file needs transcription, (2) user asks to tran...

Security Scan

The skill's files, declared requirements, and runtime instructions are coherent with a local CPU-only speech-to-text tool — it downloads a prebuilt qwen-asr binary and a model into the user's home and runs local transcription, and it does not request unrelated credentials or system access.

Purpose & Capability

Name/description (local Qwen3-ASR CPU transcription) match the declared requirement (qwen-asr binary) and the included scripts. No unrelated environment variables, credentials, or unusual binaries are requested.

Instruction Scope

SKILL.md and transcribe.sh only describe running the qwen-asr binary (locally) and converting audio via ffmpeg when needed. The install script downloads a release and the model and writes them under the user's home; scripts do not read unrelated system files or exfiltrate data. The only optional env var is QWEN_ASR_MODEL_DIR to override the model path.

Install Mechanism

Install script fetches a prebuilt release from the project's GitHub Releases and extracts it to ~/.local/bin; model download is performed by qwen-asr (presumably fetching from HuggingFace). Using GitHub Releases and the model download command is expected for this purpose; no obscure/shortened URLs or third-party personal servers are used.

Credentials

No secrets or extra environment variables are required. The only environment interaction is an optional QWEN_ASR_MODEL_DIR and use of PATH/ffmpeg. Note: model download requires network access and some HuggingFace-hosted models may require authentication in other contexts, but no credential is requested by this skill.

Persistence & Privilege

Skill is not forcible (always:false) and does not modify other skills or system-wide agent settings. It writes binaries/models into the user's home directories (standard for local tools) but does not request elevated privileges.

Assessment

This skill appears consistent with its stated purpose. Before installing: (1) review the install script and the GitHub release it downloads from and only run it if you trust that repo (prebuilt native binaries execute with your user privileges); (2) expect ~1.5 GB model download and network access to GitHub/HuggingFace; (3) ffmpeg is required for non‑WAV inputs; (4) install writes to ~/.local/bin and ~/.openclaw/tools/qwen-asr — you may need to add ~/.local/bin to your PATH; (5) if you need fully air-gapped/local operation, verify the model is cached locally or that the model download will not require a HuggingFace token. Otherwise the skill is internally coherent and proportionate.

Local, CPU-only speech-to-text powered by Qwen3-ASR. No API key or cloud needed.

• Source code: huanglizhuo/QwenASR
• Based on: antirez/qwen-asr (original C implementation)

Install

Run the install script to download the pre-built binary and model:

bash {baseDir}/scripts/install.sh

This will:

1. Download the qwen-asr binary for your platform from GitHub Releases
2. Download the qwen3-asr-0.6b model (~1.5 GB) from HuggingFace

Usage

Transcribe an audio file

bash {baseDir}/scripts/transcribe.sh <audio-file>

Supports any audio format: wav, mp3, m4a, ogg, flac, opus, webm, aac, etc. Non-WAV files are automatically converted via ffmpeg (must be installed).

Or call qwen-asr directly (WAV only):

qwen-asr -d ~/.openclaw/tools/qwen-asr/qwen3-asr-0.6b -i <audio-file> --silent

From stdin

cat audio.wav | qwen-asr -d ~/.openclaw/tools/qwen-asr/qwen3-asr-0.6b --stdin --silent

Common parameters

Flag	Description
--silent	Print only transcription text (no progress)
--language <lang>	Force language (e.g., zh, en)
-S <seconds>	Segmented mode — split audio into chunks
--stream	Streaming mode — process audio in real time
--stdin	Read audio from stdin

Model path

Default model directory: ~/.openclaw/tools/qwen-asr/qwen3-asr-0.6b