Echostrata, for Obsidian

Record. Transcribe. Locally.
Record your microphone and your Mac's system audio directly within the note you're writing.
Get a transcript, a summary and action items. Nothing leaves your Mac.

---

Why Echostrata

Meeting tools upload your conversations to someone else's servers. Echostrata doesn't.

• Mic + system audio. Captures your voice and what your Mac plays: Zoom, Teams, Meet, a browser tab, a video.
• Directly within Obsidian. Click record in a note. The audio, transcript and summary land in that same note.
• Knows who spoke. Your mic and the system audio are recorded on separate channels, so lines are labelled Me and Others without any speaker-recognition model.
• Summaries that fit the conversation. A local model writes a title, summary and to-do checkboxes. Six built-in summary templates cover meetings, 1:1s, client calls, lectures, user interviews and voice memos — or define your own custom templates with a name, purpose and prompt.
• Nothing leaves your Mac. Transcription runs on whisper.cpp, summaries on Ollama. No account, no API key, no subscription, no cloud.

What you get in your note

![[Recording 2026-10-20 10.31.05.m4a]]

> [!abstract] Q4 roadmap sync
> The team agreed on the Q4 order: search first, then sharing. Beta date confirmed.
>
> **Key points**
> - Search ships before sharing
> - Beta locked for November 14
>
> **Action items**
> - [ ] Draft the launch checklist
> - [ ] Share the beta build with the team

> [!quote]- Transcript · 24:13 · Me / Others
> `00:00` **Me:** Let's lock the Q4 roadmap today.
>
> `00:07` **Others:** Search ships first, then sharing.

The transcript is folded by default so the summary stays on top.

Features

One-click recording	A mic button in the note header turns into a red ■ 03:12 timer while recording. The ribbon icon, a command and the status bar also start and stop recording.
Floating recording bar	Live waveform, separate Mic and System level meters, plus Pause, Stop and Discard buttons. After you stop, it shows transcription progress.
Per-track mute	Click the mic or speaker icon on the recording bar to instantly mute or unmute that track mid-recording. The icon turns red, the level meter fades, and the waveform contribution drops to zero — without stopping the recording. Also available as two Command Palette commands so you can assign hotkeys.
Live note updates	A "Recording…" placeholder appears where your cursor is. It becomes "Transcribing…", then the final result.
Built-in system audio	Uses Apple's ScreenCaptureKit. No virtual audio driver, no Multi-Output Device, and your volume keys keep working.
Speaker labels	Me / Others, with names you can change. Silent channels are skipped automatically.
Local summaries	Choose any model installed in Ollama from a dropdown.
Custom templates	Write your own summary template — give it a name, a "use for" description and a prompt. Custom templates appear alongside built-in ones in the template picker.
Guided setup	A setup assistant checks and installs everything without opening Terminal.
Long sessions	Audio is written to disk as it records, and your Mac is kept awake while recording and transcribing.
Clean output	Whisper artefacts such as [BLANK_AUDIO], (music) and repeated lines are filtered out.

Requirements

• macOS 13 Ventura or later for built-in system-audio capture. Older macOS versions can use BlackHole instead.
• Obsidian 1.5.7+, desktop.
• Homebrew, used to install whisper.cpp.
• Xcode Command Line Tools, used to build the small system-audio helper once. They're already installed if you have Homebrew.
• Ollama, optional, for summaries.
• About 0.6 GB of disk for the default speech model, plus 2.6 GB if you add qwen3:4b for summaries.

Installation

Manual

1. Download main.js, manifest.json and styles.css from the latest release.
2. Copy them into <your vault>/.obsidian/plugins/echostrata/.
3. In Obsidian, go to Settings → Community plugins and enable Echostrata.

With BRAT

1. Install the BRAT community plugin.
2. BRAT → Add beta plugin → paste this repository's URL.

Setup (about 5 minutes, no terminal)

The setup assistant opens the first time you enable the plugin. You can reopen it any time from Settings → Echostrata → Open setup. Each row turns green when it's ready:

Step	What it does
Microphone access	Checks that macOS lets Obsidian use the microphone.
Built-in system audio	Click Prepare to build a small local helper (one time, about 30 s).
Speech engine	Installs whisper-cpp with Homebrew. It uses the Apple Silicon GPU.
Speech model	Downloads the default model (Large v3 Turbo q5, ~570 MB, multilingual).
Summaries (Ollama)	Optional. Pick an installed model, or download qwen3:4b / qwen3:8b with one click.

First recording: macOS asks to allow Obsidian under Privacy & Security → Screen & System Audio Recording. Allow it, then quit and reopen Obsidian. The permission only takes effect after a restart.

Check it works: play a video, then click Settings → Echostrata → Test system audio. You should see ✅ and a level reading.

Usage

1. Open a note and put your cursor where the recording should go.
2. Click the 🎙 button in the note header, or run Echostrata: Start / stop recording in current note.
3. Talk, join your call, play your video. Use Pause or Discard from the floating bar if you need to.
4. Click Stop. Transcription starts right away, and the note fills in when it's done.

Processing time is roughly 1 minute per 20 minutes of audio, plus a few seconds for the summary.

Commands

Command
Start / stop recording in current note
Pause / resume recording
Discard current recording
Record with template…
Mute / unmute microphone
Mute / unmute system audio
Open setup assistant

Assign hotkeys in Settings → Hotkeys → search "Echostrata".

Summary templates

Template	Use it for	Writes
Meeting (default)	Team meetings, syncs, workshops	Summary, decisions, key points, action items, open questions
1:1	Manager / report check-ins, mentoring	Topics, wins, concerns, feedback, commitments, next check-in
Client call	Discovery, pre-sales, account reviews	Context, needs, pain points, objections, stakeholders, deal signals, next steps, follow-up email
Lecture & learning	Courses, talks, webinars, tutorials	Outline, concepts, examples, takeaways, references, review questions
User interview	User research, usability tests	Profile, goals, pains, insights with evidence, opportunities, quotes
Voice memo	Thinking aloud, ideas, drafting	Core idea, ideas, structure, tasks, open questions, tags
Custom	Whatever you define	Whatever your prompt asks for

Custom templates are created in Settings → Summary → Custom templates → Add template. Give it a name, an optional "use for" description, and a prompt. The prompt receives the transcript and must ask the model to return structured JSON — the schema is appended automatically. Custom templates appear in the Record with template… picker and the per-note frontmatter alongside the built-in ones.

The template is chosen in this order, most specific first:

1. Record with template… (command palette), for one recording.
2. The note's echostrata-template property, for notes that are always the same type:

   ---
   echostrata-template: client-call
   ---
   

   Valid values: meeting, one-on-one, client-call, lecture, user-interview, voice-memo, plus the slugified ID of any custom template you've created (e.g. custom-weekly-standup).
3. Settings → Summary template (default: Meeting).

qwen3:4b handles every template. qwen3:8b gives noticeably better Client call and User interview results. If the model returns invalid JSON, Echostrata retries once. If that fails too, the transcript is still written, without a summary.

Settings

Section	Setting	Default
Audio sources	Microphone	System default
	System audio	Built-in macOS capture (BlackHole, a specific device or Off are also available)
	Test system audio	—
	Microphone echo cancellation	On
	Microphone level / System audio level	1.0 / 1.0
Notes	Insert recording at	Cursor position
	Recordings folder	Obsidian's attachment folder
Transcription	Speech model	Large v3 Turbo (q5)
	Language	Auto-detect
	Label speakers	On
	Speaker names	Me / Others
	Show timestamps	On
	Fold transcript callout	On
	whisper.cpp binary	Auto-detect
	CPU threads	Automatic
Summary	Generate a summary	On
	Summary template	Meeting
	Custom templates	Add / edit / delete your own templates
	Ollama URL	http://127.0.0.1:11434
	Ollama model	Dropdown of your installed models
Other	Keep Mac awake	On

Speech models

Model	Size	Best for
Base	~150 MB	Quick drafts
Small	~490 MB	English, fast
Large v3 Turbo (q5)	~570 MB	Best balance, multilingual (default)
Large v3 Turbo	~1.6 GB	Highest quality

Privacy

Echostrata is built so that everything that should stay private, stays private.

• Recording happens inside Obsidian. Audio files are saved in your vault, next to your notes.
• Transcription runs on your Mac with whisper.cpp. No audio is uploaded anywhere.
• Summaries run on your Mac with Ollama, which listens on 127.0.0.1 by default. Your transcript never goes to a remote AI service.
• No telemetry. No analytics, no account, no API keys.
• Temporary files (separate mic, system and mixed WAV tracks) are written to your system's temp folder and deleted when processing finishes.

The only network traffic is the one-time setup downloads you trigger yourself:

• whisper.cpp, via Homebrew (brew.sh) — triggered when you click Install in the setup assistant;
• the speech model, from Hugging Face (huggingface.co) — triggered when you click Download in the setup assistant;
• Ollama models, from the Ollama registry (ollama.com) — triggered when you click Download in the setup assistant.

These requests go to brew.sh and ollama.com. No request is ever made without you explicitly clicking a setup button. After setup, Echostrata works fully offline.

If you point Ollama URL to a remote server, transcripts will be sent there for summarising. Keep the default 127.0.0.1 to stay fully local.

System access disclosure

Echostrata is desktop-only and, to do its job locally, goes beyond Obsidian's vault API. It uses two powerful Node.js/Electron capabilities that Obsidian's plugin review flags. Here is exactly what they are used for, and nothing else:

Filesystem access (Node.js fs, os, path)

What	Where	Why
Writes temporary WAV files (mic, system, mix)	System temp folder, obsidian-note-recorder/<id>/	Streams audio to disk so long recordings don't fill memory. Deleted when processing finishes.
Writes the helper's Swift source	System temp folder, nrt-system-audio-v*.swift	Input for the one-time swiftc build below.
Reads and writes the plugin folder	<vault>/.obsidian/plugins/echostrata/	Stores the downloaded speech model (models/) and the compiled system-audio helper (bin/).
Reads the file you choose in the Locate on this Mac… picker	A path you pick	Lets you point to an existing whisper-cli binary.

The recording itself (.m4a) and the note text are written through Obsidian's vault API, not fs.

Shell execution (Node.js child_process)

Programs are started directly with an argument list, never through a shell string, and none of this runs until you start recording, transcribing or setup. The only program you can choose yourself is the whisper.cpp binary path.

Command	When	Purpose
whisper-cli / whisper-cpp (auto-detected, or the path you set in whisper.cpp binary)	After you stop a recording	Local transcription.
brew install whisper-cpp	Only when you click Install in setup	Installs the speech engine.
xcode-select -p / --install	Setup assistant	Checks for, or installs, the Command Line Tools.
swiftc	Only when you click Prepare in setup	Compiles the bundled Swift source into the ScreenCaptureKit helper on your Mac.
Compiled system-audio helper	While recording with built-in capture	Streams system audio to the plugin over a pipe.
caffeinate -i -m	While recording or transcribing (if Keep Mac awake is on)	Prevents idle sleep.
open / open -a	Setup assistant buttons	Opens System Settings or Audio MIDI Setup.
osascript	Only when you click Install BlackHole	Opens Terminal with brew install --cask blackhole-2ch.

Network

External domains contacted by Echostrata:

Domain	When	What
brew.sh	You click Install in the setup assistant	Downloads whisper-cpp via Homebrew
ollama.com	You click Download for a model in the setup assistant	Downloads an Ollama model
huggingface.co	You click Download for a speech model in the setup assistant	Downloads a Whisper model file

All three are user-initiated one-time actions. No request happens automatically or in the background.

Runtime: the plugin makes HTTP requests to the Ollama URL you configure (default 127.0.0.1 — local only) to list and query installed models. If you change the URL to a remote host, requests to that host will include your transcripts.

Release integrity

Release assets (main.js, manifest.json, styles.css) are built by GitHub Actions from a tagged commit with npm ci, and carry GitHub build-provenance attestations. Verify a download with gh attestation verify main.js --repo SeneKela/echostrata-for-obsidian.

How it works

Microphone ──────────┐
                     ├─► Web Audio graph ──► MediaRecorder ──► Recording.m4a (your vault)
System audio ────────┘         │
(ScreenCaptureKit helper       └─► AudioWorklet: 48 kHz → 16 kHz, level meters
 or BlackHole)                          │
                                        └─► mic.wav · system.wav · mix.wav (temp, streamed to disk)

Stop ─► whisper.cpp on mic.wav + system.wav ─► merged by timestamp ─► Me / Others paragraphs
     ─► Ollama (JSON, chosen template) ─► structured summary ─► written into the note

• System audio: a tiny Swift program using Apple's ScreenCaptureKit captures what the Mac plays and streams it to the plugin. It's compiled on your Mac from source bundled in the plugin, so there's no prebuilt binary to trust.
• Note updates: the in-note placeholder is marked with invisible Obsidian comments (%%…%%), so it's updated in place and the markers are removed once the final content is written.

Troubleshooting

Symptom	Fix
"Microphone access denied"	System Settings → Privacy & Security → Microphone → enable Obsidian, then restart Obsidian.
Only Me in the transcript	System audio is blocked or Off. Allow Obsidian under Screen & System Audio Recording, restart, then run Test system audio.
Test system audio heard silence	Make sure something is actually playing. If it is, it's almost always the permission above.
"Building the system-audio helper failed"	Install the Command Line Tools (xcode-select --install), then click Prepare again. Or switch System audio to BlackHole.
"Transcription is not set up yet"	Open the setup assistant and finish Speech engine and Speech model.
No summary appears	Ollama isn't running, or no model is selected. Open Ollama, then pick a model in Settings → Ollama model.
Using BlackHole and you can't hear anything	Your output is set to BlackHole alone. Use a Multi-Output Device (speakers + BlackHole) in Audio MIDI Setup.

Building from source

git clone https://github.com/SeneKela/echostrata-for-obsidian
cd echostrata-for-obsidian
npm install
npm run build   # → main.js
npm run dev     # watch mode

Then copy main.js, manifest.json and styles.css into your vault's plugin folder.

Project layout:

src/
  main.ts           plugin entry, commands, recording lifecycle
  recorder.ts       Web Audio graph, AudioWorklet, MediaRecorder
  native-audio.ts   ScreenCaptureKit helper (Swift source, build, stream)
  devices.ts        microphone / loopback device detection
  session.ts        recording session + processing job (save → transcribe → summarise)
  transcriber.ts    whisper.cpp runner, models, paragraphs
  summarizer.ts     Ollama models and chat requests
  templates.ts      summary templates: prompts and note rendering
  template-modal.ts "Record with template…" picker
  note.ts           in-note block insert / update / remove
  ui.ts             header button, floating bar, ribbon, status bar
  setup.ts          setup assistant
  settings.ts       settings tab
  resampler.ts      streaming 48 → 16 kHz resampler
  wav.ts            streaming WAV writer
styles.css          Figma-style UI

Roadmap

• Custom summary prompts and templates
• Live transcript while recording

Credits

• Inspired by Lecture Transcriber by AbyssLotus.
• Speech recognition: whisper.cpp by Georgi Gerganov, using OpenAI's Whisper models.
• Local language models: Ollama.
• Optional loopback driver: BlackHole by Existential Audio.

License

MIT © Martin Burkel