Local Meeting Transcriber

An Obsidian plugin that turns meeting recordings into structured notes — entirely on your machine. No cloud, no subscriptions.

Pipeline: Audio file → whisper.cpp (speech-to-text) → Any local or remote LLM (note generation) → Obsidian note

---

Features

• 🎙️ Transcribe any audio file (m4a, mp3, wav, ogg, flac, mp4, webm, aac)
• 🤖 Compatible with any OpenAI-compatible LLM server — local or remote
• 📝 Generates structured notes: title, participants, summary, discussion, action items, decisions
• 🔧 Fully configurable — system prompt, output folder, frontmatter, ffmpeg path
• ⚡ Auto-start your local server from within Obsidian

---

Privacy and Security

• The plugin is desktop-only and uses local system binaries such as whisper-cli and ffmpeg.
• It reads the audio file you choose and writes the generated note into your vault.
• It can connect to the Whisper health-check URL and the LLM API URL you configure.
• If you point the LLM URL at a remote service, your transcript and prompt data are sent to that service.
• It can run user-configured shell commands to start local servers, but only if you explicitly enter those commands in settings and trigger them manually or enable auto-start.
• The plugin itself does not require an account, does not include ads, and does not include telemetry.

---

Requirements

• Obsidian desktop (macOS, Windows, Linux)

• whisper.cpp for transcription:

  brew install whisper-cpp   # macOS
  

  Build from source for other platforms.

• A whisper GGML model (download from huggingface.co/ggerganov/whisper.cpp):

  mkdir -p ~/.whisper-models
  curl -L https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin \
       -o ~/.whisper-models/ggml-base.en.bin
  

• An OpenAI-compatible LLM server — pick one:

  Backend	Format	Install
  ollama	GGUF	brew install ollama
  llama.cpp	GGUF	build from source
  vmlx-serve	MLX (Apple Silicon)	install vMLX.app
  mlx_lm.server	MLX (Apple Silicon)	pip install mlx-lm
  LM Studio	GGUF + MLX	download app
  OpenAI / Groq / Anthropic proxy	—	set API key in settings

• ffmpeg for m4a/mp4/webm conversion (most common meeting formats):

  brew install ffmpeg   # macOS
  

  Leave the ffmpeg path empty in settings to auto-detect.

---

Installation

From Obsidian Community Plugins (recommended)

Settings → Community plugins → Browse → search "Local Meeting Transcriber" → Install → Enable

Manual installation

1. Download main.js, manifest.json, styles.css from the latest release
2. Place in .obsidian/plugins/local-meeting-transcriber/ inside your vault
3. Enable in Settings → Community plugins

---

Quick Start with ollama

1. Install ollama: brew install ollama
2. Pull a model: ollama pull llama3
3. Configure the plugin:
   • whisper-cli path: /opt/homebrew/bin/whisper-cli
   • Whisper model path: ~/.whisper-models/ggml-base.en.bin
   • LLM URL: http://127.0.0.1:11434
   • Model name: llama3
   • LLM start command: ollama serve
   • Enable Auto-start LLM server
4. Click the microphone icon in the ribbon → drop your audio file → Start pipeline

---

Configuration Reference

Speech-to-Text

Setting	Default	Description
whisper-cli path	(auto-detect)	Path to whisper-cli. Auto-detects from /opt/homebrew/bin and /usr/local/bin
Whisper model path	—	Path to a .bin GGML model file
Default language	en	Language code or auto for auto-detection
Whisper server URL	http://127.0.0.1:8178	Used for health checks only
Auto-start Whisper	off	Run the start command on plugin load
Whisper start command	—	e.g. whisper-server --model ~/.whisper-models/ggml-base.en.bin

Audio Conversion

Setting	Default	Description
ffmpeg path	(auto-detect)	Leave empty to auto-detect. Required for m4a, mp4, webm, aac

LLM Server

Setting	Default	Description
LLM URL	http://127.0.0.1:11434	Base URL of any OpenAI-compatible server
Model name	—	Model ID (e.g. llama3, gpt-4o, mistral)
API key	—	Bearer token. Leave empty for local servers
Auto-start LLM	off	Run start command on plugin load
LLM start command	—	e.g. ollama serve or vmlx-serve serve /path/to/model --port 8000

Prompt

Setting	Default	Description
System prompt	(built-in)	Editable. Must instruct the LLM to return JSON with the required fields
Speaker hint	—	Injected as context into every request

Output

Setting	Default	Description
Meetings folder	Meetings	Vault-relative folder where notes are saved
Include raw transcript	on	Append collapsible full transcript to the note

---

Output Format

Each meeting note is saved as YYYY-MM-DD Title.md with:

---
tags:
  - meeting
  - [content tags from LLM]
created: 2026-04-20
type: meeting-note
note_type: general
participants:
  - Alice
  - Bob
duration_minutes: 45
audio_file: "recording.m4a"
audio_transcribed: true
---

# Quarterly Planning Session

> [!summary]
> Two-sentence executive summary.

## Discussion

### Topic One
Flowing narrative…

### Topic Two
…

## Action Items

- [ ] Task description [Alice] [by 2026-04-27]

## Decisions

- Decided to adopt new framework by Q3.

---

## Raw Transcript

<details>
<summary>Expand raw transcript</summary>

Full whisper output…

</details>

• type: meeting-note enables Dataview queries: FROM "Meetings" WHERE type = "meeting-note"
• - [ ] action items surface in the Tasks plugin
• <details> keeps the note clean while preserving the full searchable transcript

---

Troubleshooting

whisper-cli not found → Run brew install whisper-cpp or set the path explicitly in settings.

ffmpeg not found → Run brew install ffmpeg or set the ffmpeg path in settings.

LLM server offline → Start your server manually, or set a start command in settings and enable Auto-start.

"model name not configured" → Set the model name in Settings → LLM Server.

m4a transcription fails → Ensure ffmpeg is installed. The plugin auto-converts m4a to WAV before passing to whisper.

LLM returns malformed JSON → The plugin falls back gracefully — the raw response appears in the Discussion field. Try a larger/better model or adjust the system prompt.

Transcription takes too long → Use a smaller whisper model (ggml-tiny.en or ggml-base.en). The medium and large models are much slower.

---

Development

git clone https://github.com/Zhiy-ox/Local_Voice_Transcriber.git
cd Local_Voice_Transcriber
npm install
npm run dev   # watch mode
npm run build # production build

Copy main.js, manifest.json, styles.css to your vault's plugin folder to test.

---

License

MIT © Zhiyu Xu