EchoEtcher

Voice memos to structured Obsidian notes — local Whisper transcription piped through a local LLM for formatting, tagging, and organization.

I do most of my thinking while driving - working through problems, coming up with project ideas, planning next steps. The ideas are there, but capturing them in a usable form is hard, and scrubbing through a raw voice recording later isn’t much better.

EchoEtcher watches a folder for audio files, transcribes them with Whisper running locally, processes the transcript through an LLM, and drops a clean, tagged Markdown note into Obsidian.

What it does

Record a voice memo - a thought, a project idea, a problem to work through. EchoEtcher picks it up automatically, transcribes it locally using Whisper, then sends the transcript through Ollama (or optionally Claude) to extract structure: headings, key points, tags, and clean prose. The result lands in Obsidian as a properly formatted note.

For longer recordings, I chunk the audio with overlap to keep context across segments. A custom Whisper prompt guides transcription toward natural, conversational speech - proper punctuation and sentence structure rather than raw dictation output. It handles conversational speech well, though trailing off mid-thought still trips it up occasionally.

How it works

flowchart LR
  A[Voice memo] --> B[File watcher]
  B --> C[Whisper transcription]
  C --> D[LLM formatting]
  D --> E[Obsidian note]

Python with a file watcher (watchdog) monitoring a configurable directory. Whisper runs locally - it auto-detects CUDA, MPS (Apple Silicon), or falls back to CPU. The processing pipeline supports multiple content types through a modular processor system, so audio and images can be handled differently.

It runs as a background agent with configurable polling intervals, designed to run unattended and process files as they appear. Everything stays local unless the Claude formatting step is opted into.

Updates

2026-04-26

Currently offline while I work on some convenience improvements to the workflow. Was in regular use before that.