skills.rest
xiaopengde
xiaopengdeCOMMUNITY
Explore all skills by this author →
GitHub

murmur

Transcribe local audio recordings into cleaned Markdown or Word documents using Whisper.

110|12|Updated May 14, 2026
Search Tags:#offline#speech recognition#transcription#whisper#meeting notes#docx#audio-to-text
One-click install
npx skills add https://github.com/xiaopengde/murmur --skill murmur-xiaopengde
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: murmur
Source: https://github.com/xiaopengde/murmur
Command: npx skills add https://github.com/xiaopengde/murmur --skill murmur-xiaopengde

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires mlx-whisper, whisper-ctranslate2, and includes scripts (resource) and references (resource) components.

What problem does it solve? Paid cloud transcription services like Feishu Minutes, Tongyi Tingwu, or Otter.ai charge per minute and upload your recordings to the cloud. This Skill runs the full meeting/interview transcription pipeline entirely on your own machine: local Whisper inference, LLM-based cleanup into a readable document, and optional Word export, with zero cost and zero data leaving your device. ## Core Features & Use Cases - Local Whisper transcription: Converts m4a/mp3/wav/mp4 audio to text using large-v3-turbo via mlx-whisper on Apple Silicon or whisper-ctranslate2 on Windows/Linux/Intel Mac, with ffmpeg preprocessing to avoid hallucination loops. - LLM cleanup workflow: Assembles a cleaning prompt (via scripts/clean.py) that fixes ASR errors, labels speakers, and structures the transcript into chapters, then optionally converts the result to docx with pandoc. - Environment automation: doctor scripts check dependencies, install scripts set up ffmpeg/uv/pandoc on macOS, Windows, and Linux, and mainland-China network detection routes model downloads through ModelScope or mirrors. - Use Case: Drop a 30-minute interview recording into your agent and say "transcribe this" — the Skill checks the environment, runs onboarding once (default format and model), transcribes offline, and delivers a cleaned Markdown or Word document. ## Quick Start Use Murmur to transcribe the file meeting.m4a on my Desktop into a cleaned meeting document.

Frequently Asked Questions about murmur

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I transcribe an audio file to text locally without a cloud service?▼

Run the transcribe.py script with your audio file after the install script sets up ffmpeg, uv, and pandoc. It converts the audio to 16kHz WAV with ffmpeg, runs Whisper large-v3-turbo locally, and outputs a raw transcript plus an SRT subtitle file.

What is the best free alternative to Otter.ai or Feishu Minutes for meeting transcription?▼

A local Whisper-based pipeline replaces per-minute SaaS transcription at zero cost with full privacy. This Skill uses large-v3-turbo for accuracy, then an LLM cleaning prompt fixes ASR errors, labels speakers, and structures the output into a readable document.

Does local Whisper transcription work on Windows without a GPU?▼

Yes, Windows and Linux use whisper-ctranslate2, which runs on CPU at roughly 1-2x audio duration. An NVIDIA GPU with CUDA speeds it up automatically, and you can switch to a smaller model like medium for faster CPU runs.

Why does Whisper transcription repeat the same phrase or hallucinate text?▼

Repetition loops usually come from condition-on-previous-text being enabled or from feeding compressed audio directly. The transcribe script disables that flag by default and always converts input to 16kHz mono WAV with ffmpeg first.

Can Whisper transcription identify different speakers in a recording?▼

Whisper itself has no speaker diarization, so speakers are inferred by the LLM during the cleanup step from conversational context. For chaotic multi-speaker audio, the output falls back to generic Speaker 1/2/3 placeholders.

How do I download Whisper models faster in mainland China?▼

The script auto-detects mainland China by timezone and locale, then downloads verified models from ModelScope and routes package installs through the Tsinghua PyPI mirror. You can force this with the --cn flag or persist it via --set-default-cn on.

Related Skills

Explore More Content & Communication Skills→
junghan0611junghan0611
6

transcribe

Transcribe audio files to plain text via the Groq Whisper API.

Community
Intermediate
alexandereiseghohialexandereiseghohi

meeting-minutes

Generates structured meeting summaries with owners and tracked actions from text input.

Community
Intermediate
peteanderson80peteanderson80

transcribe

Transcribe audio and video to text with optional speaker diarization.

Community
Intermediate
astoreyaiastoreyai
3

stt-transcription

Transcribe audio from files or microphones using multiple speech-to-text engines.

Community
Advanced