A picture frame that listens to the room and paints what you talk about. Speech-to-text and topic spotting run locally, each topic becomes a cartoon, and the cartoons fill a collage on a colour e-ink panel, a TV or any screen.
An example page as the settings page previews it, laid out and dithered by E-maki's own code for a 13.3" Spectra 6 panel. The topics are made up, and the pictures are hand-drawn stand-ins for the "Ink and red" style.
Warning
Early and experimental. The whole pipeline runs end to end in Docker on a laptop, but no complete frame has been built yet. Expect rough edges and breaking changes.
Warning
The project is vibe-coded and should be treated as such.
Claude Code wrote most of the code. I steered the design, tested it on my own laptop and decided what to keep. The guidelines it works under are in CLAUDE.md, including the privacy rules every change has to follow.
I'm a hobbyist, and I fell for Fugleramme: an e-ink frame that shows a collage of the birds BirdNET-Go hears in the garden. I wanted the same idea for the people in the room, a frame that slowly fills with what we have been talking about. The topics a small local model picks out of a dinner conversation are often funny, and that is half the point.
An emaki (絵巻) is a Japanese illustrated handscroll that tells a story in pictures as you unroll it, which is roughly what the frame does with a conversation. The e (絵) means picture, and it also stands for e-ink.
- Listen. A mic records the room in 15-30 second segments. Audio stays in memory and is never saved.
- Transcribe. whisper.cpp turns each segment into text on your own machine. It can send Danish and Norwegian to fine-tuned models that you point it at (Languages).
- Spot the topic. About once a minute a local LLM in Ollama reads the last few minutes and names the topic in English, as specifically as it can ("Is a hot dog a sandwich?", not "food").
- Paint it. A new topic becomes a short image prompt, which is sent to an image API with your own key. This prompt is the only thing about your conversations that leaves your network. Or draw the pictures locally with stable-diffusion.cpp, and nothing leaves at all.
- Show it. The picture joins the collage. Topics that keep coming up grow, one-offs fade. The page is dithered for a Spectra 6 e-ink panel and served over HTTP.
Two modes: Continuous paints a topic as soon as it comes up, for when guests are over. Diary buys no pictures during the day. It collects the topics that came up at least twice, and once a day (diary_at, 07:00 by default) paints the top few, so you wake up to yesterday on the wall and the frame refreshes once a day. Diary is meant for a frame on the wall. A new install starts in Continuous mode so you see results on the first run. Switch under Frame > Timing on the settings page.
Tip
E-maki is a single Go binary. Speech-to-text and the LLM run as separate containers it talks to over HTTP, and how it reaches them lives in a small config profile. It runs on a laptop today, with integrated GPU support through Vulkan. The page is a plain PNG, so any browser can show it.
E-maki is an always-on microphone, so:
- Audio lives only in RAM and is never written to disk.
- Transcripts live only in an in-memory rolling window. They are never written to disk or logged, and they stay on your network unless you switch on an external topic model (below). Logs show sound levels, languages and word counts, not words.
- Speech-to-text and the topic LLM run on your own machine, on a Docker network with no route to the internet. Their models are downloaded by separate one-shot containers that never see audio or text.
- With the local topic model, the only thing about your conversations that leaves the house is the short English image prompt, sent with your own key. Without a key nothing does. (The first start downloads the models, but sends nothing about you.) A small local LLM writes the prompt, and code checks it before it goes out (no numbers, no quotes from the transcript, a length limit). It can still slip in a first name or a place now and then. Switch on Spoken names under Listening > Privacy to also refuse prompts that repeat a name it heard.
- Optional, off by default: an external topic model (OpenAI, Gemini, OpenRouter, Groq or any OpenAI-compatible chat API) instead of the local one, for when your machine is too slow for a local LLM. This sends the transcript window, what was said in the last few minutes, to that service. Set
TOPIC_API_KEYandTOPIC_API_MODELin.envand pick it on the settings page under Listening > Topics, which shows a warning while it is on. - Sensitive topics (health, money, relationships, confidential work) are blocked before any image request, under Listening > Privacy. A blocked topic gets no image and no tile, and its label is not logged. All four are blocked by default. The local LLM makes the call, so it will sometimes get it wrong.
- The page itself shows what was talked about, and any device on your local network can open it. Set a password under System > Security; see Lock it down.
Tell the people in the room that it is listening. PRIVACY.md lists everything E-maki keeps, logs and sends.
Note
Only the laptop setup has been run so far. The frame parts below are the ones Fugleramme uses, but no E-maki frame has been built yet, and the dithering and .e6 output are only checked against Fugleramme's in tests. If you try it, please share how it went in Discussions.
| Part | What works | Notes |
|---|---|---|
| Brain | A Linux computer that runs Docker: a laptop, a mini PC, a homelab box. Windows works through WSL | An integrated GPU makes speech-to-text about 7x faster. A Jetson Orin Nano Super is the plan for a small build, untested |
| Mic | The brain's own mic or a USB mic | Or a Pi or an ESP32 that streams over Wi-Fi |
| AI image provider | Your own Gemini, OpenAI or OpenRouter key, or a local model | Optional: without one you get coloured placeholder tiles |
| Screen | Any browser to start with | For the real thing, a 13.3" Spectra 6 e-ink panel (Inky Impression) on a Raspberry Pi, or a battery-powered ESP32 board, in an IKEA RÖDALM frame |
Parts, wiring and flashing in Hardware. Installing, .env, the settings page, the frame and logs in Setup.
To try it on a laptop:
git clone https://github.com/Strandgaard96/E-maki && cd E-maki
cp .env.example .env # optional: add your own GEMINI_API_KEY
docker compose up --buildThen open http://localhost:8090/ and talk for a minute.
To show it on a Raspberry Pi with an Inky Impression, install Fugleramme's Inky client on the Pi:
curl -fsSL https://raw.githubusercontent.com/arnegiacomo/fugleramme/main/examples/inky/install.sh \
| FRAME_URL=http://<your-brain>:8090 bashEvery setting lives in one place, so the two never disagree:
- The settings page (
http://<your-brain>:8090/): everything about what the frame shows, listening, topics, privacy and which image provider draws. Hover the ? next to a setting to see what it does. .env(copy from .env.example): API keys, the Whisper and topic models, the GPU and other extras, the mic socket. Change it, then rundocker compose up -d. System > Services shows what it is set to. See Setup.
How E-maki reaches its servers lives in a small profile in configs/. With docker you don't need to touch it.
Needs Go (version in go.mod) and a C compiler such as gcc, for the miniaudio library behind the mic capture.
go run ./cmd/emaki -script testdata/party.txt # offline: fake LLM and images, synthetic transcript
go run ./cmd/emaki -config configs/laptop.json -mic # with whisper-server on :8081 and Ollama on :11434
go run ./cmd/emaki -config configs/laptop.json -wav talk.wav # the same, from a 16 kHz mono WAVThe settings page is at http://localhost:8090/, the picture at http://localhost:8090/frame.png. Add -settings settings.json to keep what the page changes. The code is laid out like this:
| Package | Role |
|---|---|
internal/audio |
microphone capture, network microphone, WAV files, level check |
internal/stt |
whisper-server client, per-language routing |
internal/topic |
topic LLM over an OpenAI-compatible API, privacy check, offline fake |
internal/memory |
topic memory: decay, "seen twice", daily cap, pin and remove |
internal/imagegen |
Image providers (Gemini, OpenAI-style APIs including the local sd-server, OpenRouter) and an offline placeholder |
internal/collage |
The page at the panel's size: one big tile plus up to 8 smaller tiles with captions |
internal/panel |
Fugleramme dithering and .e6 (ported), /frame.png, /frame.e6, /press |
internal/timeline |
the saved pictures behind Pictures > History |
internal/web |
the settings page (Preact in TypeScript, Tailwind and Basecoat; source in internal/web/ui, built files embedded) |
internal/pipeline |
wires it together |
Issues and PRs are welcome.
- Something is broken: a bug report
- A question, an idea, or a frame you have built: Discussions
- A fix or a doc change: open a PR, no issue needed
- A security or privacy problem: see the security policy, not a public issue
Read the privacy rules in Contributing first.
- Fugleramme by Arne Giacomo Munthe-Kaas: the bird frame E-maki grew out of. Its Spectra 6 dithering,
.e6format and panel protocol are ported to Go ininternal/panel. - BirdNET-Go by tphakala: the always-on local listener that classifies what it hears, and the idea of running on whatever hardware you have, integrated GPUs included.
- whisper.cpp and Ollama, which do the speech and language work, and stable-diffusion.cpp for local pictures.
E-maki is not affiliated with any of them.
- Code: MIT. See
LICENSE. - Preact (MIT), Tailwind CSS (MIT) and Basecoat (MIT) for the settings page, and Flickr's justified-layout (MIT) and PhotoSwipe (MIT) for its Timeline.
- Code ported from Fugleramme (
internal/panel): MIT, Copyright (c) 2026 Arne Giacomo Munthe-Kaas. Seethird_party/fugleramme/LICENSE. - Sidecars (whisper.cpp, Ollama, stable-diffusion.cpp): MIT. Built or pulled as containers, not bundled.
- Models (Whisper and Qwen by default, plus Edda, NB-Whisper, Gemma or FLUX.2 klein if you pick them): downloaded on first start and covered by their own licences. Check each model card before you use one.
- Generated images: made with your own API key and covered by your image provider's terms, or made locally and covered by the model's licence.