Transcriber and Summary for Idiots

Records and lecture audio in, finished notes out. For each recording you get a title, a summary, key points, and the full transcript — written into the note that already links the recording.

Everything runs on your own computer. No accounts, no API keys, no subscription, and nothing is ever uploaded.

Setup, step by step

You do not need to use a terminal. Follow these in order.

1. Install the plugin

In Obsidian: Settings → Community plugins → Browse → search Transcriber and Summary for Idiots → Install → Enable.

Not listed yet? Use BRAT: install the BRAT plugin the same way, then BRAT → Add beta plugin → paste AbyssLotus/obsidian-lecture-transcriber.

2. Let it install the speech parts

The first time it loads, a Setup window appears and lists what is missing. Press Install what is missing and wait. It downloads about 1.1 GB.

If the window is gone, reopen it: Settings → Transcriber and Summary for Idiots → Set up.

On a Mac there is no ready-made Whisper download, so the plugin uses Homebrew. If you do not have it, the window tells you: install Homebrew from brew.sh, then press Install what is missing again.

You can stop here. Transcription now works. Steps 3 and 4 only add the summary and key points.

3. Install Ollama (only for summaries)

Summaries are written by a second program called Ollama. It is free and also runs entirely on your computer.

1. Download it from ollama.com/download
2. Install it like any other app
3. Open it once. It has no real window — it just runs quietly in the background. That is normal and correct.
4. Back in Obsidian, open the Setup window again

4. Let it download a summary model

With Ollama running, the Setup window offers two buttons. Press one:

A progress bar shows the download. When it finishes, summaries are on.

Using it

Press the microphone in the left sidebar. It transcribes every recording in your vault that has not been done yet, showing progress per file.

Or just record straight into Obsidian and forget about it — new recordings are transcribed on their own a few seconds after they appear.

Where the notes go. If a note embeds the recording (![[Recording 1.m4a]]), the transcript goes into that note. Otherwise a note beside the recording with the same name is used, or created.

How long it takes. Roughly 1 minute of processing per 20 minutes of audio, plus about 20 seconds for the summary. A one-hour lecture takes about 4 minutes. Your computer is kept awake for the whole thing, including while recording.

Study notes from several lectures

Separate from everything above, and it never runs on its own — only when you ask for it.

Transcripts are long and you will not reread them. This takes a set of lectures and writes one set of notes that actually explains the material: an overview of the topic, the key concepts, and a vocabulary table.

To use it: command palette → Generate study notes from lectures, or Settings → Study notes → Write study notes.

Pick the course folder, then choose which lectures to include. There are shortcuts for Today, This week, Last two weeks and Everything, and you can tick individual lectures. Dates come from the lecture note names, so grouping by week works even if the files were edited later.

It writes a new note beside the lectures and never touches them. Each one links back to the lectures it came from.

Roughly half a minute per lecture: three lectures of a calculus course, about 15,000 words of transcript, took 80 seconds.

It is allowed to explain beyond the transcript. The model may use its own knowledge of the subject to define terms properly and explain ideas the lecturer rushed, because a garbled definition is worse than a textbook one. It is forbidden from inventing course specifics — deadlines, assignments, what is on the exam — unless the transcript says so. You can turn this off in settings to keep strictly to what was said.

If something goes wrong

"Not set up yet" — open Settings → Transcriber and Summary for Idiots → Set up and press the button.

Transcripts appear but no summary — Ollama is not running, or has no model. Open Settings → Test the summary model → Test. It tells you exactly what is wrong in one line. Usually: open the Ollama app, then press Test again.

Windows: "the program did not start" or "a required DLL is missing" — the Whisper program needs Microsoft's C++ runtime, which many Windows installs do not have. Install Microsoft Visual C++ Redistributable (x64) from aka.ms/vs/17/release/vc_redist.x64.exe, restart Obsidian, then press Install again. This is the single most common Windows setup failure.

Windows: it downloads but will not run, and the runtime is already installed — antivirus or Windows itself is usually blocking a newly downloaded program. Open the folder the error names, double-click whisper-cli.exe once, and allow it if Windows asks. Then press Install again.

It sits at 0% for a long time — on Windows and Linux there is no graphics acceleration available, so transcription runs entirely on the processor and is genuinely slow: roughly ten to thirty minutes per hour of audio on a laptop. Whisper also reports nothing until the first chunk finishes, so 0% is normal at the start. The elapsed clock beside each recording confirms it is still working, and an estimate of the time remaining appears once it passes 5%. To make it faster, pick a smaller model under Summary and key points → Whisper model.

Everything is slow — Settings → Summary model: pick a smaller one. qwen3:4b is several times faster than a large model.

It keeps mishearing the same word — Settings → Corrections, one rule per line:

gravel => parabola
Dr Thurman => Dr Thierman

Those are applied to every transcript from then on, so the fix is permanent.

The transcript has a warning box on it — the recording confused Whisper badly enough that some of the lecture is probably missing. Usually very quiet or very low-quality audio. Record closer to the speaker if you can.

The Mac still sleeps with the lid shut — nothing can prevent that; macOS does not allow it. Leave the lid open.

Settings worth knowing

• Corrections — permanent find-and-replace for words it keeps getting wrong.
• Extra vocabulary — names, course codes and jargon to expect.
• Use my notes to improve accuracy — on by default. What you have already typed in the note makes those words far likelier to be recognised.
• Write maths as LaTeX — on by default, so $x = -\frac{b}{2a}$ renders as a real equation.
• Also mark up maths in the transcript — off by default, and slow (about 35 minutes for a one-hour lecture). Worth it for maths classes only.
• Keep the screen on too — leave this on. When the display sleeps, Obsidian gets throttled and a summary in progress can fail.
• Let the model think first — off by default. About three times slower for no real gain on lecture material.

What it accesses, and why

Obsidian's directory flags this plugin for filesystem access and for running programs. Both are real and unavoidable: speech recognition is a separate program that reads and writes files. Here is the whole of it.

It has no third-party dependencies. package-lock.json lists zero external packages, so the only code here is the code in this repository. main.js in each release is a byte-for-byte copy of src/main.js — node scripts/build.js prints the hashes, and releases carry a signed GitHub build-provenance attestation you can check:

gh attestation verify main.js --repo AbyssLotus/obsidian-lecture-transcriber

Files it reads and writes

It never reads or writes vault files outside Obsidian's API, and it never touches files elsewhere on your computer.

Programs it runs

Network requests

Nothing else is contacted. Your recordings, transcripts and notes are never sent anywhere — transcription and summarising both run locally, which is the reason the downloads are large.

What it does not do

• Mobile. Desktop only. It runs real programs on your computer, which Obsidian on phones and tablets cannot do.
• Identify who is speaking. The transcript is one continuous voice.
• Fix bad recordings. A phone across a lecture hall will transcribe poorly no matter the settings.

Notes on accuracy, if you care

Measured on real lecture recordings while building this:

• The full large-v3 model family is clearly better than turbo on lecture speech — turbo rendered "algebraic equation" as "algebraic queen".
• A denoiser made things worse, not better: it caused duplicated sentences. None is used.
• Text-context carry-over is disabled deliberately. With it on, a 97-minute low-bitrate recording fell into a repetition loop 30 minutes in and never recovered, losing two thirds of the lecture. Turning it off costs nothing measurable in accuracy.
• Audio is decoded and levelled inside Obsidian, which is why ffmpeg is not needed. This was checked to transcribe identically to the ffmpeg version.

For developers

node test/run-all.js                                   # run the suites
node scripts/build.js                                  # assemble dist/
VAULT=/path/to/vault node scripts/build.js --install   # install into a vault

The suites run against a stub of the Obsidian API, so CI exercises the platform-specific code on Windows, macOS and Linux without needing Obsidian.

Licence

MIT