菜单
One-time setup for local Whisper transcription — installs faster-whisper and downloads a default model. Use when the user asks to set up whisper, install whisper, or prepare for local transcription.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Export a transcript (or any pipeline output) by emailing it, uploading to Google Drive, or copying to clipboard. Use when the user asks to send, email, upload, save to Drive, or copy a transcript.
Transcribe a podcast episode and produce two outputs in one pass — (1) the raw API transcript exactly as the provider returned it, and (2) a podcast-formatted reading version with filler words removed, section headers added, and speaker labels preserved. Use whenever the user provides a podcast audio file or says "transcribe this podcast".
Diagnose-and-treat workflow for very messy recordings where the standard preprocessor isn't enough. Generates a spectrogram, identifies problematic frequency bands (AC hum, rumble, narrow-band whines, broadband hiss, clipping), and proposes targeted ffmpeg filters. ONLY use when the user explicitly invokes it — the normal pipeline runs `preprocess-for-transcription` directly and skips diagnostic work. Trigger phrases the user might use — "this audio is messy", "transcription failed because of noise", "fix this rough recording", "the audio is bad and the transcript is garbage".
Remove background noise from an audio file. Cloud providers (Auphonic default, ElevenLabs, Dolby.io) or local (DeepFilterNet ML, ffmpeg afftdn non-ML). Use when the user asks to denoise, clean up noise, remove hum/hiss, or enhance speech audio before transcription.
Prepare audio for transcription — format normalise (mono/16kHz/opus), loudness normalise (EBU R128), and collapse long silences (silero-vad). Optional denoise pass. Use before sending to AssemblyAI or any ASR for cleaner results, smaller uploads, and lower cost. Use when the user asks to "preprocess audio", "prep for transcription", "clean up a recording before sending it", or just hands over a raw voice memo to be transcribed.
Apply a stylistic / formatting transformation to a transcript using a named prompt from the user's Text-Transformation-Prompt-Library (206 transformations covering blog outlines, briefs, business correspondence, analysis documents, meeting notes, and more). Use when the user asks to "convert this to <style>", "format like a <type>", or names a specific transformation. Fetches the catalog on demand and caches it; never commits the prompts into this plugin.
| name | setup-whisper |
| description | One-time setup for local Whisper transcription — installs faster-whisper and downloads a default model. Use when the user asks to set up whisper, install whisper, or prepare for local transcription. |
One-time install for offline transcription.
Ensure uv is available (which uv). If not, install: curl -LsSf https://astral.sh/uv/install.sh | sh.
Install faster-whisper as a tool:
uv tool install faster-whisper
(Or create a dedicated venv if the user prefers — check preference.)
Pre-download the default model (medium) so first real transcription doesn't block:
uv run --with faster-whisper python -c "from faster_whisper import WhisperModel; WhisperModel('medium')"
Write/update ~/.config/claude-transcription/config.json with:
{ "whisper_installed": true, "whisper_model": "medium", "whisper_device": "cpu" }
Confirm with a 5-second synthetic test (optional) — or just tell the user setup is complete.
Ask the user which model they want pre-downloaded:
base — 150 MB, fastestsmall — 500 MB, good for most usemedium — 1.5 GB, default balancelarge-v3 — 3 GB, highest accuracyIf the user wants ROCm acceleration, warn that CTranslate2 ROCm support is experimental and document the extra steps (install ctranslate2 built against ROCm, set whisper_device: "cuda"). Default to CPU.