Openai Whisper
clawhubVideo Editing Medium
3.9
OpenAI Whisper：使用 Whisper 模型进行视频语音识别和转录。
815
clawhub install openai-whisper
Local speech-to-text with MLX Whisper (Apple Silicon optimized, no API key).
# Install Skill (downloads SKILL.md to .claude/skills/) clawhub install mlx-whisper # Then just tell Claude: "use Mlx Whisper to help me..."
# Same install command — works with all SKILL.md-compatible AI coding tools clawhub install mlx-whisper
This Skill is compatible with the OpenClaw standard. After installation, a SKILL.md file is auto-generated, usable by any OpenClaw-compatible AI Agent (Claude Code, Cursor, Windsurf, etc.).
mlx_whisper /path/to/audio.mp3 --model mlx-community/whisper-large-v3-turbo
# Transcribe to text file
mlx_whisper audio.m4a -f txt -o ./output
# Transcribe with language hint
mlx_whisper audio.mp3 --language en --model mlx-community/whisper-large-v3-turbo
# Generate subtitles (SRT)
mlx_whisper video.mp4 -f srt -o ./subs
# Translate to English
mlx_whisper foreign.mp3 --task translate
~/.cache/huggingface/mlx-community/whisper-tiny; use --model mlx-community/whisper-large-v3-turbo for best results```bash
# 转录中文播客，输出带时间戳的 VTT 字幕
mlx-whisper podcast.m4a \
--language zh \
--model small \
--output-format vtt \
-o podcast_subtitles.vtt
# 批量转录多个音频文件
for file in *.mp3; do
mlx-whisper "$file" --language zh --model base -o "${file%.mp3}.txt"
done
```Real cold-start runs of this skill on Claude Code and Codex.
Input: Transcribe a short English speech clip (speech.wav, 16kHz mono) to text using Whisper (mlx-whisper preferred on Apple Silicon).
Ran mlx-whisper directly (MLX / Metal) — whisper-small-mlx; cross-checked tiny/base/small (agree). Only friction: HF download via SOCKS proxy needed httpx[socks]/socksio installed first.
claude-opus-4-8
MLX BLOCKED — no Metal device in Codex’s sandboxed session, so mlx-whisper could not run. Fell back to CPU openai-whisper (base.en); model downloaded fine. Same transcript.
codex-cli 0.140
Output: Transcript (of the ~2.5s TTS clip): “Welcome to skills.video, the…” — both agents agree; Whisper accurately rendered the audible content (the clip itself trails off mid-sentence; the TTS truncated the source).
Verdict: Both transcribed accurately. Key agent-environment finding: Codex’s sandbox has no Metal/GPU access → it can’t run mlx-whisper and falls back to CPU openai-whisper; Claude Code (unsandboxed) uses MLX directly. Mirrors the Chromium-sandbox finding — Codex’s sandbox restricts hardware/process access that Claude’s does not.
Environment: macOS (Apple Silicon) · symmetric cold start. Test audio: a ~2.5s English TTS clip (macOS `say` truncated a longer sentence — a harness limit, not the skill). HF model downloads route via a SOCKS proxy → httpx[socks] required first.