Funasr Asr

本地中文语音识别，使用阿里达摩院 FunASR。触发场景：(1) 用户发送语音消息需转录，(2) 需要转录音频/视频文件，(3) 从网页下载视频并转录，(4) 中文语音转文字任务。支持小内存模式（~500MB）和大模型模式（~2GB），自动内存管理，任务队列防并行。

Security Scan

The skill's files, runtime instructions, and requirements are coherent with a local FunASR-based speech-to-text tool; nothing in the bundle asks for unrelated credentials or surprising system access.

Purpose & Capability

The name/description (local FunASR ASR) match the code and docs: Python scripts load FunASR models, transcribe audio, extract audio from video, and optionally download videos. Dependencies (funasr, onnxruntime, yt-dlp/ffmpeg) and model downloads are proportional to the stated purpose.

Instruction Scope

SKILL.md and code instruct running local Python scripts, installing Python packages, and optionally downloading videos via yt-dlp or requests. The scripts read predictable local paths (/tmp locks, ~/.cache/modelscope), inspect memory (/proc/meminfo or psutil), and do not attempt to read arbitrary unrelated system configuration or secrets. Note: processing a video URL will fetch external web content (user-supplied) and invoke yt-dlp/ffmpeg, which is expected for the feature.

Install Mechanism

No registry install spec was provided, but package.json runs scripts/install.sh which pip-installs funasr and onnxruntime. There are no downloads from untrusted shorteners or personal IP addresses; model acquisition is expected to occur via modelscope/ModelScope tooling or the funasr library. This is a standard pip-based install (traceable but subject to normal PyPI supply-chain risks).

Credentials

The skill declares no required environment variables or credentials. Documentation mentions optional MODELSCOPE_CACHE and common OS tools but does not require unrelated secrets. File reads/writes are limited to model cache and temporary working directories.

Persistence & Privilege

The skill does not request always:true, does not modify other skills, and only creates per-skill temporary files and locks under /tmp; it does not try to persist or escalate privileges beyond its own runtime files.

Assessment

This skill appears to be what it claims: a local FunASR-based transcriber. Before installing, consider the following: (1) it requires Python and pip and will install funasr and onnxruntime (PyPI packages) — standard but subject to supply-chain risks; (2) first run will download model files (~500MB to ~2GB) into your model cache (~~/.cache/modelscope/), so ensure disk and bandwidth; (3) video transcription will fetch arbitrary web URLs (yt-dlp/requests/ffmpeg) — avoid supplying internal/private URLs you don't want the runtime to contact; (4) scripts spawn subprocesses (ffmpeg, yt-dlp, python) and write temporary files under /tmp and a model cache — run in an environment with appropriate resource limits or isolation if you are concerned; (5) review or pin the exact Python packages you install if you need stronger supply-chain guarantees. If you want, ask for a list of the exact network endpoints used for model downloads or for an explanation of the expected files the skill will create and where (model cache, temp dirs, lock files).

FunASR 本地语音识别

阿里达摩院开源 ASR，完全本地部署，内存优化版。

快速使用

命令行

# 默认 small 模式（SenseVoiceSmall, ~500MB）
python3 scripts/transcribe.py /path/to/audio.wav

# 大模型模式（Paraformer-Large, ~2GB，中文极高精度）
python3 scripts/transcribe.py /path/to/audio.wav --mode large

# 自定义分段时长（默认 600 秒 = 10 分钟）
python3 scripts/transcribe.py /path/to/audio.wav --segment 300

# JSON 输出
python3 scripts/transcribe.py /path/to/audio.wav --format json

# 转录视频文件
python3 scripts/video-transcribe.py --audio /path/to/video.mp4

Node.js 调用

const funasr = require('./index.js');

const text = await funasr.transcribe('/path/to/audio.wav', {
  mode: 'small',    // 'small' 或 'large'
  format: 'text'    // 'text' 或 'json'
});

核心优化（v2.0）

优化项	旧方案	新方案
进程模型	每段新进程	单进程，模型加载一次
默认模型	paraformer (~2GB)	SenseVoiceSmall (~500MB)
内存峰值	~2GB × N 次	~500MB 常驻
段间释放	进程退出	gc.collect()
任务锁	✅	✅ 保留

模式对比

特性	small (默认)	large
模型	SenseVoiceSmall	Paraformer-Large
内存	~500MB	~2GB
语言	中英日韩粤	中文
精度	高	极高
速度	快 (~0.1 RTF)	较慢 (~0.3 RTF)
适用场景	日常转录	专业中文场景

安装

# Python 依赖
pip install funasr onnxruntime psutil

# 视频处理（可选）
pip install yt-dlp
apt install ffmpeg

# 首次运行自动下载模型
# small: ~500MB | large: ~2GB

支持格式

• 音频: WAV, MP3, FLAC, M4A（自动转 16kHz mono）
• 视频: MP4, WebM, AVI, MOV

故障排查

问题	解决
模型下载失败	pip install modelscope && modelscope download --model iic/SenseVoiceSmall
内存不足 (OOM)	用 --mode small --segment 300 减小分段
音频格式错误	ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav
转录有奇怪标记	脚本自动清理 `<

许可证

MIT License