EchoNote: say it once, find it organized

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend

What I Built

I built EchoNote, a small local voice-to-action desktop app.

I built it for myself and my friends who get ideas, reminders or things to do but don't want to stop what they are doing, open a notes app and start typing.

The idea is very simple.

Press Alt + Shift + Space, speak whatever you want, press it again and continue working. (You can configure another key combo as well)

EchoNote takes what you said and automatically organizes it into a task, reminder, idea or note.

For example, you can say:

"Remind me to submit the hacktoberfest weekend challenge at 10 tomorrow morning."

And EchoNote will create a reminder for tomorrow.

The important part is that everything runs locally on the laptop. Your voice and notes don't go to a server.

I have not shown it to my friends yet. I spent this weekend mainly making the capture system reliable and testing it. For now, I am the first user.

Demo

EchoNote is available here if you want to give it a try:

Stable release:
https://github.com/KrAG2000/echonote/releases/tag/v1.0.5

(Setup takes just 5 minutes, except for the model download, which depends on your internet speed.)

The release page has builds and installation instructions for the supported systems: an .rpm for Fedora, a .deb for Debian/Ubuntu, and an AppImage for everything else.

Capture: press the shortcut or the button, speak, and the note is filed automatically.

Reminders: "Remind me tomorrow… to call the dentist" became a reminder for tomorrow. EchoNote shows the exact words it used for the date.

First run: the models are downloaded once and checked. After that, nothing leaves your computer.

Privacy: EchoNote runs locally on your machine. Your voice, notes and AI processing stay on your computer, and after the initial model download, the app does not need an internet connection.

Code

GitHub: https://github.com/KrAG2000/echonote

The project is open source under the MIT license.

How I Built It

I built EchoNote using Electron, React and TypeScript.

For speech-to-text I used whisper.cpp.

For understanding and organizing the voice input, I used Google's Gemma 4 E2B running locally through llama.cpp.

The app uses SQLite to save the notes.

The models are downloaded once during setup. After that, EchoNote does not need the internet to work.

I also tested the app without a network connection to make sure the offline part actually works.

I tested different Whisper and LLM models and selected them based on speed and how correctly they organized the sentences.

I also built the app so that the original transcript is saved before the AI processes it. So even if the AI is slow or not available, the note is not lost.

Why Does Open Innovation Matter?

For this project, open models made it possible to build something that is private and runs completely on my laptop.

These are personal notes, ideas and reminders. I don't want to send them to some server just to organize them.

There is also no API key or per-request cost.

I could test different models on my own laptop and choose what worked best for the app.

It also means the app can work without internet.

EchoNote currently runs entirely on the CPU, so it works on laptops without a GPU. GPU acceleration is on my list for the next version.

The downside is that local AI can be slower on CPU compared to a hosted AI API. But for me, the privacy and offline use are worth it.

Also, Gemma 4 e2b, the model is super lightweight and I learnt that it is built for on device efficient processing. Since a lot of users use mobile normally, my next goal is to release an android version. Recent android versions come with a lot of restrictions, will have to figure out "how". AI is there to help ;)

My Agent Session

I built EchoNote with Claude Code. Honestly, I have no idea how linux apps are made, claude made this app possible. I just had a few ideas. I actually created a whole another app for this submission before I came to EchoNote. Previous app felt irrelevant and useless tbh. Wasting 2 of my precious days!

Agent session:

Building EchoNote: a local voice-to-action assistant with whisper.cpp + Gemma 4 (highlights, updated)

You

Project Specification: EchoNote — Local-First Voice-to-Action Desktop Assistant

Document type: Product requirements, system architecture, implementation plan, and acceptance criteria
Project status: New project — Hacktoberfest Weekend Challenge 2026
Target platform for the first release: Linux desktop, developed and tested on Fedora
Target user: A friend or loved one who frequently has thoughts, tasks, ideas, and reminders that they forget to record or organize
Primary implementation stack: Electron, React, TypeScript, Node.js, SQLite, local speech recognition, and local open-weight LLM inference
Primary constraint: Deliver a complete, reliable, installable first version before expanding the feature set.

---

1. Instructions to the Coding Agent

You are responsible for implementing this application end to end inside the current project directory.

Do not merely produce a plan, mockup, architecture diagram, or a collection of disconnected components. Build the actual application, install its dependencies, run it, test the critical workflows, fix implementation errors, and prepare a working build.

1.1 How to approach the work

1. Inspect the existing project directory before modifying anything.
2. Identify the operating system, available development tools, repository state, and existing code.
3. Read the entire specification before making architectural decisions.
4. Create an implementation plan and then begin implementation without waiting for additional instructions.
5. Prefer the simplest architecture that satisfies the requirements.
6. Use actual local speech recognition and actual local language-model inference. Do not replace AI processing with hardcoded sample outputs or simulated results.
7. Do not require the end user to install Node.js, Python, Ollama, a compiler, a database server, or a separate model runtime.
8. Keep the application usable when the LLM is unavailable or still loading.
9. Test the complete application rather than testing only individual functions.
10. Document any requirement that cannot be completed and explain the reason. Do not silently omit it or claim it works without verification.

1.2 Implementation priorities

When time or technical constraints force trade-offs, use this order:

1. Reliable recording and persistent capture.
2. Fast, accurate local speech-to-text.
3. Local classification into tasks, reminders, ideas, and reference notes.
4. Persistent storage and a usable inbox.

[… the remaining 25 sections of the specification are omitted here …]

Create this app end to end!

Show more

Agent

The app starts in about 230 ms, but the global shortcut registration returned false under Wayland. I'll investigate which backend Electron is using.

Agent

Still false. The portal session may set itself up asynchronously, so I'll retry the registration after a delay.

Agent

Now I'll make the shortcut manager try the Electron global shortcut first and fall back to the GNOME desktop binding.

Agent

Both phases are unusually slow for a 1.5B model on this CPU: prompt processing about 55 tok/s, generation about 12 tok/s. Thread count barely matters. I'll check whether the machine is under load.

Agent

Orphans are cleaned up. I'm now fixing two issues: making inference servers die with the app, and a fast shortcut path (about 10 ms instead of about 2.7 s) through a private named pipe.

Agent

I've built EchoNote end to end and verified it on your Fedora machine. You record with a shortcut and whisper.cpp transcribes locally. Qwen2.5-1.5B (via llama.cpp) files the note as a task, reminder, idea or reference note, and reminders fire as desktop notifications. The installable build is dist/EchoNote-1.0.0-x86_64.AppImage (135 MB), and everything is in one local git commit (c61bb47, branch main, not pushed).

What I tested
- Unit tests: 96 pass (dates, validation, database, pipeline, reminders, downloads, input checks).
- Real-model tests: 8 pass, running the actual speech and AI models with nothing mocked.
- End-to-end tests: 7 workflows driving the built app, with WAV files fed in as the microphone and recording toggled exactly the way the keyboard shortcut does it:
- A reference note is saved and still findable after a restart.
- A "remind me in two minutes" reminder fires once and doesn't fire again after a restart.
- A future reminder is still scheduled after a restart.
- With the AI model missing, the note waits in the Inbox, then gets organized once the model is back, without duplicates.
- A failed model download, then a successful real download.
- A denied microphone, an invalid shortcut, and rapid repeated toggles, which create only one recording.
- A full capture with no network interface at all, to prove it works offline.
- I re-ran the key workflows against the packaged app and the AppImage itself.
- One open issue: the offline test failed once in six runs, and I couldn't reproduce it or capture the error.
- Not tested: a real voice (test audio is robotic synthetic speech), a physical key press on the GNOME shortcut, and Hindi/Hinglish.

Performance on this laptop (CPU only, power profile "quiet")

Step	Time
Window appears	0.2–0.7 s
Key press reaches the app	17–27 ms (1.4 s if EchoNote isn't running yet)
Text visible after you stop talking	~2.3 s
Note fully organized	~7.5–8.3 s
AI step alone	~5.4 s, 8/8 correct on my test sentences
Memory	AI ~1.9 GB, speech ~0.1–0.16 GB, app ~0.7–1 GB

The spec's target of AI classification under 2 s is not met on this hardware. The smaller 0.5B model takes about 2.8 s but got 6/8 right, so it's offered as the low-memory option. The text is saved first, so nothing waits on the AI. GPU acceleration isn't implemented because there's no CUDA or Vulkan SDK on this machine. Switching GNOME to Balanced or Performance mode should help.

Things that work differently from the spec
- Shortcut on GNOME/Wayland: Electron can't register global shortcuts there (tested). Instead, Settings → Set up GNOME shortcut adds a GNOME custom keyboard shortcut that signals the running app. I tested install and removal against your real GNOME settings and they restore exactly, but I didn't press the key.
- Models aren't inside the AppImage: first launch downloads 1.1 GB from Hugging Face and checks it; after that the app never uses the network.
- Reminders only fire while EchoNote is running. Closing the window keeps it running in the background.
- The tray icon won't show on stock GNOME without the AppIndicator extension.

Bugs the testing caught and I fixed:
- A broken GNOME shortcut write.
- AI servers left running if the app was killed hard.
- A 2.7 s shortcut delay, now 17–27 ms.
- A race that made the first note after startup take 18 s.
- A first recording that failed when the shortcut itself launched the app.

What's left for you before the deadline (Oct 5, 12:29 PM IST)
1. Record a few real notes with your own voice and press the actual GNOME shortcut.
2. Give it to your friend and note what they say. The draft has placeholders for this; please don't fill them with anything that didn't happen.
3. Record a 60–90 s demo video.
4. Create a public GitHub repo, push, and attach the AppImage to a release.
5. Copy docs/DEV_SUBMISSION_DRAFT.md into the official template from the challenge announcement post (the challenge data I could fetch doesn't include it). Add the repo and video links, keep the tags (#hf26challenge, #devchallenge, #weekendchallenge), and publish.

Your disk has about 1.3 GB free and the models use 1.6 GB of it. I didn't delete anything of yours. If you don't want the low-memory model, you can remove ~/.config/EchoNote/models/qwen2.5-0.5b* to recover 0.5 GB.

Files are in ~/Personal_Projects/h26_dc_personal_assistant:
- README.md
- docs/DEV_SUBMISSION_DRAFT.md
- docs/benchmarks/README.md
- dist/EchoNote-1.0.0-x86_64.AppImage

Show more

You

Create git repo and push it with proper commit message(s).

Agent

The code is now on GitHub as a public repo: https://github.com/KrAG2000/echonote. main there matches your local copy exactly.

Before pushing, I replaced the single local commit with 9 commits, one per area of the project, each with a descriptive message. The files are identical to what was there before. Each commit ends with the Claude co-author line. The commits are:

69612cd docs: README, architecture, privacy, troubleshooting, benchmarks
517cbf7 test: unit, real-inference integration and end-to-end workflows
fa337d8 feat(ui): capture, inbox, category views, search, settings and setup
2d945d4 feat(app): Electron shell, secure IPC, overlay, tray and shortcuts
3d2a29f feat(pipeline): local transcription, classification and reminders
a445182 feat(db): SQLite storage with migrations and repositories
0251618 feat: shared types, zod schemas and typed IPC contract
ed934c8 build: native whisper.cpp/llama.cpp runtimes, model manifest and icons
0e2d953 chore: scaffold Electron + React + TypeScript project

A few things are not in the repo:
- The AppImage and the compiled whisper/llama binaries. These are build outputs and are excluded on purpose. npm run native:build rebuilds the binaries.
- The model files. The app downloads them on first run.
- docs/DEV_SUBMISSION_DRAFT.md is public now. It still has the [FRIEND'S NAME] placeholders. Fill them in or delete the file before you share the repo link.

For the challenge, people need something to download. Should I create a v1.0.0 GitHub release with EchoNote-1.0.0-x86_64.AppImage and its SHA256SUMS attached? That would make the README's install instructions work.

Show more

You

Change draft to real version. I havenot shown it to any friend, so say so. Also, any placeholders to fill in, make them real. If you have any questions, feel free to ask me.

Agent

The post draft is finished and pushed (commit 6a7b314): https://github.com/KrAG2000/echonote/blob/main/docs/DEV_SUBMISSION_DRAFT.md

I also created the release, so the post links to a real download: https://github.com/KrAG2000/echonote/releases/tag/v1.0.0. It contains the AppImage and its checksum file, and I checked the checksum against the file.

What changed in the draft:
- Your text stays. Your title, your intro in your own words and your planned-features paragraph are all kept, with only typos fixed ("commmuting", "occured", "many a times"). I added one line saying who it's for, based on your answer: you and your friends, creative people whose ideas are lost if they aren't captured right away.
- Friend: a short paragraph says plainly that you haven't shown it to any friend yet, and that you'll give it to them next.
- Demo: three real screenshots (all links checked and working), plus a line saying the video is coming soon.
- Code: your ![EchoNote](…) line was image syntax, which shows as a broken image on DEV. I replaced it with ordinary links to the repo and the release, plus install commands.
- How I Built It: you had commented out the heading together with the diagram. I left the diagram out but restored the heading, because the bullet points under it need one.
- Everything else is filled in: the "Why open matters" and "Where open was harder" sections, the bugs testing caught, and the honest performance numbers. The "What [FRIEND] Said" section is gone; your planned features are now listed under What's Next.
- I added a one-line disclosure at the end that it was built with an AI coding agent, which DEV's rules expect. Edit the wording if you like, but keep it.

Before you publish:
1. Template: paste the content into the official template from the challenge announcement post. I couldn't get the template itself, only the rules.
2. Video: record the demo video and replace the "coming soon" line with the link.
3. Your own voice: the post says the shortcut and recording work. Try it with your own voice once before publishing, since all my tests used synthetic audio.
4. Deadline: publish by Oct 5, 12:29 PM IST with the tags #hf26challenge #devchallenge #weekendchallenge #opensource.

I can also stage it as an unpublished draft on your DEV account if you want; you'd still review it and press publish yourself.

Show more

You

Fully local. this project do not comes under any prize category. Its local 100%. All prize category is cloud hosted. We can be in gemma categroy if we use gemma in this project. How is it?

Agent

Gemma fits well, because it's an open-weight model that runs fully locally through llama.cpp, the runtime we already ship. I'll check which Gemma builds are available, their size, and whether downloading them requires a Hugging Face login, before recommending anything.

Agent

Gemma 4 E2B with thinking switched off, then with a compact output grammar:

▸ Bash S=<scratch>; B=resources/bin/linux-x64/llama-server; M=~/.config/EchoNote/mod...

Input

$ S=<scratch>; B=resources/bin/linux-x64/llama-server; M=~/.config/EchoNote/models/gemma-4-E2B-it-Q4_0.gguf; export PATH=~/.nvm/versions/node/v24.21.0/bin:$PATH; $B --help 2>&1 | grep -iE "reasoning|think" | head -5; $B -m $M -c 2048 -ub 256 -np 1 -t 8 --no-webui --jinja --reasoning-budget 0 --host 127.0.0.1 --port 18802 > $S/g4b.log 2>&1 & PID=$!; until curl -sf localhost:18802/health >/dev/null; do sleep 0.3; done; CE=$S/ce.mjs SC=$S/sc.mjs node --input-type=module -e "
const {buildMessages}=await import(process.env.CE);const {LLM_JSON_SCHEMA}=await import(process.env.SC);
const cases=[['warm',''],['The staging server uses port 8081, not 8080.','reference'],['Remind me tomorrow at 7 pm to call the dentist.','reminder'],['I need to fix the login bug on the settings page.','task'],['I could build a tool that explains database query plans in plain English.','idea'],['My passport number is stored in the blue folder in the top drawer.','reference'],['Buy milk and eggs on the way home.','task'],['What if the app could summarize my week every Sunday.','idea'],['Pay the

Output

--reasoning-format FORMAT               controls whether thought tags are allowed and/or extracted from the
                                        - deepseek: puts thoughts in `message.reasoning_content`
                                        - deepseek-legacy: keeps `<think>` tags in `message.content` while
                                        also populating `message.reasoning_content`
                                        (env: LLAMA_ARG_THINK)
39891 ms gen 100 tok reasoning 0 exp  got task
14592 ms gen 94 tok reasoning 0 exp reference got reference
14311 ms gen 96 tok reasoning 0 exp reminder got reminder
13753 ms gen 92 tok reasoning 0 exp task got task
13669 ms gen 91 tok reasoning 0 exp idea got idea
13525 ms gen 88 tok reasoning 0 exp reference got reference
13751 ms gen 93 tok reasoning 0 exp task got task
15847 ms gen 107 tok reasoning 0 exp idea got idea
13989 ms gen 95 tok reasoning 0 exp reminder got reminder
median 13989 accuracy 8/8

You chose Gemma 4 E2B as the default. First I'll try forcing compact JSON with a hand-written grammar that allows no whitespace and fixes the key order, so the model spends tokens only on content.

▸ Bash S=<scratch>; cat > $S/compact.gbnf <<'EOF' root ::= "{\"category\":" cat ",\"...

Input

$ S=<scratch>; cat > $S/compact.gbnf <<'EOF'
root ::= "{\"category\":" cat ",\"title\":" title ",\"summary\":" summary ",\"action\":" action ",\"date_expression\":" dexpr ",\"needs_confirmation\":" bool ",\"reason\":" reason ",\"confidence\":" conf "}"
cat ::= "\"task\"" | "\"reminder\"" | "\"idea\"" | "\"reference\""
title ::= "\"" char{1,80} "\""
summary ::= "\"" char{0,400} "\""
action ::= "null" | "\"" char{1,200} "\""
dexpr ::= "null" | "\"" char{1,80} "\""
reason ::= "null" | "\"" char{1,200} "\""
bool ::= "true" | "false"
conf ::= "0" ("." [0-9] [0-9]?)? | "1" (".0")?
char ::= [^"\\\x7F\x00-\x1F] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
EOF
B=resources/bin/linux-x64/llama-server; export PATH=~/.nvm/versions/node/v24.21.0/bin:$PATH
for M in gemma-4-E2B-it-Q4_0.gguf qwen2.5-1.5b-instruct-q4_k_m.gguf; do $B -m ~/.config/EchoNote/models/$M -c 2048 -ub 256 -np 1 -t 8 --no-webui --jinja --host 127.0.0.1 --port 18803 > /dev/null 2>&1 & PID=$!; until curl -sf localhost:18803/health >/dev/null; do sleep 0.3; done; echo "== $M"; G=$S

Output

== gemma-4-E2B-it-Q4_0.gguf
  36312 ms gen 66 tok  -> task
  9397 ms gen 61 tok reference -> reference
  9933 ms gen 62 tok reminder -> reminder
  8763 ms gen 58 tok task -> task
  8208 ms gen 54 tok idea -> idea
  8708 ms gen 56 tok reference -> reference
  8764 ms gen 59 tok task -> task
  8408 ms gen 54 tok idea -> idea
  9869 ms gen 66 tok reminder -> reminder
  median 8764 accuracy 8/8
  sample: {"category":"reminder","title":"Pay electricity bill","summary":"Pay the electricity bill on the 15th.","action":"Pay the electricity bill on the 15th","date_expression":"the 15th","needs_confirmation":false,"reason":null,"confidence":0.98}
== qwen2.5-1.5b-instruct-q4_k_m.gguf
  15928 ms gen 48 tok  -> task
  6593 ms gen 67 tok reference -> reference
  5308 ms gen 55 tok reminder -> reminder
  5504 ms gen 58 tok task -> task
  5140 ms gen 53 tok idea -> idea
  5201 ms gen 53 tok reference -> reference
  5560 ms gen 59 tok task -> task
  4507 ms gen 47 tok idea -> idea
  5870 ms gen 61 tok reminder -> reminder
  median 5504 accuracy 8/8
  sample: {"category":"reminder","title":"Pay electricity bill","summary":"Pay the electricity bill on the 15th","action":"Pay the electricity bill on the 15th","date_expression":"15th","needs_confirmation":false,"reason":null,"confidence":0.95}

Agent

Gemma 4 E2B is now EchoNote's default AI model, still 100% local, and it's pushed and released. The release is https://github.com/KrAG2000/echonote/releases/tag/v1.0.1 (marked Latest). The post draft now enters Best Use of Gemma and explains why I chose it.

Why Gemma 4 E2B: I tested four models on your laptop with the same 8 test sentences:

Model	Correct	Time per note	Notes
Gemma 4 E2B (new default)	8/8	~8.6 s	Apache-2.0 license, ~4 GB RAM
Qwen2.5 1.5B (old default)	8/8	~5.5 s	stays in Settings as the faster option
Gemma 3 1B	6/8	~4.2 s	low-memory option
Qwen2.5 0.5B	6/8	~2.8 s	low-memory option

Two changes made Gemma usable on your CPU:
- Thinking off: Gemma 4 has a "thinking" mode that used the whole answer budget on hidden reasoning (over 40 s per note). It's now switched off for this task.
- Compact output: Gemma was writing indented JSON, which doubled the number of words it had to generate. The app now forces compact JSON, which cut its time from ~14 s to ~8.6 s with the same accuracy.
- Overall: from when you stop talking, the transcript still appears in about 2.3 s and the note is filed in about 11–13 s, versus 7–8 s with Qwen.

Testing: all 96 unit tests and 8 real-model tests pass with Gemma, and all 7 end-to-end workflows pass. The new AppImage passed the shortcut, missing-AI and offline tests.

I changed one test. With Gemma, the robotic test voice came out as "remind me in two minutes, just trash…". Gemma correctly asked for confirmation instead of guessing, so the reminder test now clicks Confirm the way you would. The old test required the reminder to be scheduled with no confirmation step.

You still have EchoNote 1.0.0 running, which you launched yourself, and I didn't touch it. When you switch to 1.0.1 it will use Gemma by default and show the AI as not installed. Download Gemma once (2.65 GB) in Settings → AI model; until then, new notes are transcribed and wait in the Inbox. To stay on Qwen, select it there instead.

The deadline is today, Oct 5, at 12:29 PM IST. Your remaining step is the same as before: add the video link where the draft's Demo section says the video is coming soon, then paste the draft into DEV and publish. Your demo will be most accurate on v1.0.1 with Gemma downloaded, since that's what the post describes.

Show more

You

Shoukld I not see the app listed inn the software app in linux?

Agent

No, you shouldn't expect to see it there, and nothing is broken. GNOME Software only lists apps installed through a package system: Fedora's RPM repositories (dnf) or Flatpak. An AppImage deliberately skips that and runs as a standalone file, so the system doesn't know it's "installed". The menu entry you created just points at the file.

To make EchoNote appear in Software with a normal Uninstall button, there are three options:

Option	Shows in GNOME Software?	Effort	Notes
RPM package (echonote-1.0.1.x86_64.rpm)	✅ under Installed, with Uninstall	~30–45 min to add and test	Fedora's native format. rpmbuild is already on your machine. Installing needs your password (sudo dnf install ./echonote-1.0.1.x86_64.rpm), and it adds a proper app-menu entry and icon automatically. Removing it from Software or with sudo dnf remove echonote also removes the app. Your notes in ~/.config/EchoNote stay unless you use Delete all data first.
Flatpak	✅	several hours	The most "app store"-like option, but its sandbox complicates microphone access, the bundled AI servers and the GNOME shortcut. Too risky before today's deadline.
Gear Lever (a Flatpak app from GNOME Software)	❌ (shows in Gear Lever instead)	0 for me	It manages AppImages for you: adds them to the menu and gives you an uninstall button. Good if you want to keep using the AppImage.

My recommendation is to add an RPM build alongside the AppImage. You're on Fedora, so it's the natural way to install and uninstall, and judges on Fedora get the familiar experience. The AppImage stays for everyone else.

The deadline is in about 10 hours (12:29 PM IST). If you want the RPM, I'll do it like this:
1. Add the rpm target to the build.
2. Build it.
3. Check it with rpm -qpl, making sure it contains the AI binaries and the desktop entry.
4. Attach it to the v1.0.1 release.
5. Add install and uninstall instructions to the README.

You'd then run sudo dnf install once to verify that it shows up in Software. I can't run sudo for you. Want me to go ahead?

Show more

You

One interesting idea. When recording start or stop, i want to show indication in the title basr above. Currently, a popup appears that shows recording. Can we have that? A red dot for recording, something else wehn recording stops that disappears after 3 seconds. Also, 5 seconds for recording timeout as well, if nothing is being spoken.

Agent

That's a nice upgrade. Before choosing an approach I need to check what your GNOME top bar supports. Stock GNOME doesn't show app icons up there; that needs the AppIndicator extension.

Agent

Two failures, so I'll inspect both instead of guessing. Workflow B's reminder came out 84 s after the capture instead of about 2 minutes, which could be a real date bug. The offline test timed out.

Agent

Both features are done, tested, and released as v1.0.3: https://github.com/KrAG2000/echonote/releases/tag/v1.0.3

What's new
- Red dot in the top bar. It works like GNOME's screen recorder: a red dot appears only while recording. When you stop, a green check shows for 3 seconds and then the icon disappears. At other times there's no icon.
- The popup only appears as a fallback. It shows only when the top bar can't display the dot. You can force it on or off in Settings → While recording.
- Silence auto-stop. Recording stops by itself after 5 seconds with no speech. You can change the time or turn it off in Settings → While recording. If nothing at all was said, nothing is saved. A recording you stop yourself is always kept, so a quiet speaker never loses a note.

Two bugs the new tests caught, both fixed
- "Remind me in two minutes" was rounded down to the minute, so it could fire up to 59 seconds early.
- Dates were worked out from when the AI finished, not from when you spoke. A note waiting in the Inbox for an hour would have had "tomorrow" or "in two minutes" shifted by that hour.

Testing: all 102 unit tests pass. All 8 end-to-end workflows pass, including a new one with silent audio: recording stopped on its own after 5 s and nothing was saved. The 8 real-model tests pass too, and the packaged RPM-layout build and the AppImage pass their checks. The release also includes your new feather icon and your popup CSS tweak, which were in the folder when I built it.

Not verified: I couldn't see the red dot myself, because your GNOME doesn't have the extension yet.

To get the red dot on your laptop
1. Install and enable the extension, then log out and back in:
bash
sudo dnf install gnome-shell-extension-appindicator
gnome-extensions enable [REDACTED]

2. Quit EchoNote (Settings → Quit EchoNote completely) and upgrade it:
bash
sudo dnf install ./dist/echonote-1.0.3.x86_64.rpm

3. Open EchoNote again. It checks for the top-bar support at startup, so it must be restarted after step 1. Press your shortcut and the red dot should appear.

It's 8:2x AM; submissions close at 12:29 PM IST. You still need to record the video and publish the post. The silence auto-stop and the red dot would make good moments to show in the demo.

Show more

You

Need a bundling system. I also need a .deb file alongside .rpm file.

Agent

I'll take "bundling system" to mean one command that builds every Linux package (AppImage, RPM and DEB), checks them, writes checksums, and can optionally publish the release. Before adding the .deb, I'll check whether our native binaries can even run on Debian/Ubuntu. They were compiled on Fedora 44, which has a very new glibc.

▸ Bash for f in resources/bin/linux-x64/*server resources/bin/linux-x64/libstdc++.so...

Input

$ for f in resources/bin/linux-x64/*server resources/bin/linux-x64/libstdc++.so.6 dist/linux-unpacked/echonote; do echo "$f: $(objdump -T $f 2>/dev/null | grep -o 'GLIBC_[0-9.]*' | sort -uV | tail -1)  $(objdump -T $f 2>/dev/null | grep -o 'GLIBCXX_[0-9.]*' | sort -uV | tail -1)"; done; ldd --version | head -1; which dpkg-deb ar fakeroot 2>&1 | head

Output

resources/bin/linux-x64/llama-server: GLIBC_2.43  GLIBCXX_3.4.32
resources/bin/linux-x64/whisper-server: GLIBC_2.43  GLIBCXX_3.4.32
resources/bin/linux-x64/libstdc++.so.6: GLIBC_2.38  GLIBCXX_3.4.35
dist/linux-unpacked/echonote: GLIBC_2.25  
ldd (GNU libc) 2.43
/usr/bin/dpkg-deb
/usr/bin/ar
/usr/bin/fakeroot

Agent

The .deb passes on Ubuntu 24.04:
- it installs with apt
- whisper transcribes the JFK clip correctly
- Gemma correctly answers "Reference"

The Ubuntu 22.04 check is running now.

Agent

▸ Bash L=<scratch>/dist.log; until grep -q "^EXIT" $L; do sleep 5; done; sed -n '/Ve...

Input

$ L=<scratch>/dist.log; until grep -q "^EXIT" $L; do sleep 5; done; sed -n '/Verifying the .deb on ubuntu:22.04/,$p' $L | grep -E "==|ERROR|FAIL|EXIT|E:" | cut -c1-200; grep -E "needs glibc|present|depends|  [0-9.]+M  " $L | cut -c1-200; cat ~/Personal_Projects/h26_dc_personal_assistant/dist/SHA256SUMS

Output

[1;34m==> Verifying the .deb on ubuntu:22.04[0m
== Ubuntu 22.04.5 LTS, glibc 2.35
== installed: echonote 1.0.4
== desktop entry + AppStream metadata present
== whisper-server starts (no missing symbols)
== llama-server starts (no missing symbols)
== whisper transcript: {"text":" And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.\n"}
== llama answer: "content":"Reference"
== PASS: .deb verified on Ubuntu 22.04.5 LTS
[1;34m==> Done[0m
EXIT 0
  whisper-server needs glibc 2.34
  llama-server   needs glibc 2.34
  rpm: all required files present
  deb: all required files present
  AppImage: all required files present
  deb depends: libgtk-3-0, libnotify4, libnss3, libxss1, libxtst6, xdg-utils, libatspi2.0-0, libuuid1, libsecret-1-0
  104M  dist/echonote_1.0.4_amd64.deb
  129M  dist/EchoNote-1.0.4-x86_64.AppImage
  92M  dist/echonote-1.0.4.x86_64.rpm
== desktop entry + AppStream metadata present
== desktop entry + AppStream metadata present
e87e0c1e9ae00c4a9fd13d2a3b707389a511ff54bd1709e87d3f6cd325fb22ef  EchoNote-1.0.4-x86_64.AppImage
d092904391f5a011f6a8b56ccb2890db31092a1fd4b1174d089ab4f77c12de5e  echonote-1.0.4.x86_64.rpm
004282b96e55dce526bc7c4e4ea7aadf19efdaf965bfb34bd09b21fa8b14d92d  echonote_1.0.4_amd64.deb

Agent

The .deb is done and released alongside the .rpm and AppImage, and there's now one command that builds and checks all three. It's released as v1.0.4: https://github.com/KrAG2000/echonote/releases/tag/v1.0.4

File	For	Size
echonote_1.0.4_amd64.deb	Debian / Ubuntu: sudo apt install ./echonote_1.0.4_amd64.deb	104 MB
echonote-1.0.4.x86_64.rpm	Fedora: sudo dnf install ./echonote-1.0.4.x86_64.rpm	92 MB
EchoNote-1.0.4-x86_64.AppImage	everything else	129 MB

The .deb behaves like the RPM: it adds an app-menu entry, appears in GNOME Software / App Center, and uninstalls with sudo apt remove echonote.

The bundling system: npm run dist:linux does everything in one go:
- refuses to package AI engines that wouldn't run on Ubuntu or Debian
- runs lint and the unit tests
- builds all three packages and checks each one contains the AI engines, the model list, the icons and the menu entry
- writes dist/SHA256SUMS

Two options:
- --verify-deb installs the .deb in clean Ubuntu 24.04 and 22.04 containers and checks that transcription and the AI actually work there.
- --release publishes to GitHub. It reads the notes from docs/release-notes/v<version>.md.

A bug this uncovered in every earlier release. The AI engines had been compiled on your Fedora 44, so they needed its very new system library (glibc 2.43). On Ubuntu 24.04 or Debian 12, the app would install but transcription and AI would not start. I now compile them inside Ubuntu 22.04 (via Docker), so they need only glibc 2.34. Releases v1.0.0–v1.0.3 still carry the Fedora-only engines; v1.0.4 is marked Latest, so new downloads get the fixed version.

Verified:
- The .deb installed with apt in clean Ubuntu 24.04 and 22.04 containers. Inside both, its whisper transcribed the JFK speech correctly and Gemma classified a note correctly as "Reference".
- On Fedora, all 8 real-model tests pass with the new engines, at the same speed and the same 8/8 accuracy. The AppImage passes the shortcut and silence-stop tests.

Not verified: Ubuntu's desktop experience itself (window, microphone, notifications), because the containers have no display.

To upgrade your installed copy, quit EchoNote, then run sudo dnf install ./dist/echonote-1.0.4.x86_64.rpm.

It's about 9:05 AM, with roughly 3h25m until the 12:29 PM deadline. I'd stop adding features now and record the video and publish.

Show more

You

I had installed devRelay. Do we have the session that I can embed?

Agent

You didn't have an EchoNote session on DEV, so I made one from the highlights of this conversation and added its embed to the post draft.

Your session: https://dev.to/agent_sessions/building-echonote-a-local-voice-to-action-assistant-with-whispercpp-gemma-4-highlights-lzr147 (ID 539, 26 messages)

What's in it:
- Real excerpts only, nothing rewritten: your requests and my actual replies, from the spec through Gemma, the RPM, the red-dot indicator and the .deb.
- Three key moments with their real tool output: the Gemma benchmark (14 s → 8.6 s, 8/8 correct), the glibc 2.43 discovery, and the .deb passing in Ubuntu 22.04.

Before uploading, I made these changes:
- ~ replaced with ~
- your email removed
- the long spec shortened to its first section
- long outputs trimmed

I scanned it first and found no tokens or keys. DEV's own filter also blanked one harmless GNOME extension ID.

It's unpublished. That's fine for your own post, because the author can always embed their own unpublished session. Open the link above to review it first.

In the post: the "My Agent Session" section of docs/DEV_SUBMISSION_DRAFT.md (pushed as c71dea6) now has a two-line intro and this tag:

{% agent_session building-echonote-a-local-voice-to-action-assistant-with-whispercpp-gemma-4-highlights-lzr147 %}

When you paste the draft into DEV, the session should show up there, interactive.

The deadline is 12:29 PM IST. What's left is the demo video and publishing the post.

Show more

You

TEst the app again. the version 1.0.4. I start wtih key combo or the icon click, modal pops up, goes away, again pops up. SOme kind of repitition is there

Agent

I'll start from your real logs rather than guessing: your installed version, plus the recorder events from your attempts.

▸ Bash rpm -q echonote; L=~/.config/EchoNote/logs/echonote.log; ls -la ~/.config/Ech...

Input

$ rpm -q echonote; L=~/.config/EchoNote/logs/echonote.log; ls -la ~/.config/EchoNote/logs/; grep -E "app: starting|recorder|capture: saved|pipeline|indicator|shortcut|toggle-pipe|setup|ui ready" $L | tail -60 | cut -c1-170

Output

echonote-1.0.4-1.x86_64
total 12
drwxr-xr-x. 1 user user   24 Oct  5 07:43 .
drwx------. 1 user user  820 Oct  5 10:21 ..
-rw-r--r--. 1 user user 4608 Oct  5 10:22 echonote.log
{"t":"2026-10-05T02:13:19.978Z","level":"info","msg":"app: starting","meta":{"version":"1.0.2","packaged":true}}
{"t":"2026-10-05T02:13:20.321Z","level":"info","msg":"toggle-pipe: listening"}
{"t":"2026-10-05T02:13:20.351Z","level":"info","msg":"app: ui ready","meta":{"ms":374}}
{"t":"2026-10-05T02:13:44.545Z","level":"info","msg":"shortcut: GNOME custom shortcut installed","meta":{"accelerator":"Alt+Shift+Space"}}
{"t":"2026-10-05T04:51:05.140Z","level":"info","msg":"recorder: start requested","meta":{"source":"cli"}}
{"t":"2026-10-05T04:51:06.915Z","level":"info","msg":"recorder: stop requested","meta":{"source":"cli"}}
{"t":"2026-10-05T04:51:07.100Z","level":"info","msg":"capture: saved audio","meta":{"id":"8dc16cf2-0935-4e6b-9d1c-54029a9be8e4","durationMs":1536}}
{"t":"2026-10-05T04:51:08.103Z","level":"info","msg":"recorder: start requested","meta":{"source":"cli"}}
{"t":"2026-10-05T04:51:09.639Z","level":"info","msg":"pipeline: transcribed","meta":{"id":"8dc16cf2-0935-4e6b-9d1c-54029a9be8e4","ms":2527,"chars":3}}
{"t":"2026-10-05T04:51:24.426Z","level":"info","msg":"pipeline: classified","meta":{"id":"8dc16cf2-0935-4e6b-9d1c-54029a9be8e4","ms":14776,"attempt":0}}
{"t":"2026-10-05T04:51:24.987Z","level":"info","msg":"recorder: stop requested","meta":{"source":"button"}}
{"t":"2026-10-05T04:51:25.187Z","level":"info","msg":"capture: saved audio","meta":{"id":"b0f6305a-7f9b-457e-af0e-288c5ee073b7","durationMs":16896}}
{"t":"2026-10-05T04:51:27.779Z","level":"info","msg":"pipeline: transcribed","meta":{"id":"b0f6305a-7f9b-457e-af0e-288c5ee073b7","ms":2576,"chars":42}}
{"t":"2026-10-05T04:51:37.896Z","level":"info","msg":"pipeline: classified","meta":{"id":"b0f6305a-7f9b-457e-af0e-288c5ee073b7","ms":10112,"attempt":0}}
{"t":"2026-10-05T04:51:50.797Z","level":"info","msg":"recorder: start requested","meta":{"source":"cli"}}
{"t":"2026-10-05T04:51:52.607Z","level":"info","msg":"recorder: stop requested","meta":{"source":"cli"}}
{"t":"2026-10-05T04:51:52.743Z","level":"info","msg":"capture: saved audio","meta":{"id":"387c0c27-6db0-4afc-9fef-073cb579c88a","durationMs":1536}}
{"t":"2026-10-05T04:51:54.871Z","level":"info","msg":"pipeline: transcribed","meta":{"id":"387c0c27-6db0-4afc-9fef-073cb579c88a","ms":2119,"ch

Agent

▸ Bash ps -eo pid,ppid,lstart,args | grep -E "/opt/EchoNote/echonote( |$)|echonote-t...

Input

$ ps -eo pid,ppid,lstart,args | grep -E "/opt/EchoNote/echonote( |$)|echonote-toggle|tee /run" | grep -v -- "--type=" | grep -v grep | cut -c1-200; echo; K=org.gnome.settings-daemon.plugins.media-keys; gsettings get $K custom-keybindings; gsettings get $K.custom-keybinding:/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/echonote/ command; gsettings get $K.custom-keybinding:/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/echonote/ binding; ls -la $XDG_RUNTIME_DIR/echonote/; python3 -c "import hashlib,os;print('expected pipe tag', hashlib.sha1(os.path.expanduser('~/.config/EchoNote').encode()).hexdigest()[:10])"

Output

3706331    4624 Mon Oct  5 07:34:35 2026 /opt/EchoNote/echonote
3725243    4624 Mon Oct  5 07:43:18 2026 /opt/EchoNote/echonote

['/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/echonote/']
"'sh' '-c' 'f=$1; shift; if [ -p \"$f\" ] && printf t | timeout 1 tee \"$f\" >/dev/null; then exit 0; fi; exec \"$@\" --toggle' 'echonote-toggle' '/run/user/1000/echonote/toggle-35be96832d' '~/Personal_Projects/h26_dc_personal_assistant/dist/EchoNote-1.0.4-x86_64.AppImage'"
'<Alt><Shift>space'
total 0
drwx------.  2 user user  60 Oct  5 09:03 .
drwx------. 17 user user 800 Oct  5 07:49 ..
prw-------.  1 user user   0 Oct  5 07:43 toggle-35b32893b9
expected pipe tag 35b32893b9

Agent

Unit tests pass. Next I'll prove the shortcut fix with a before/after check: snapshot your GNOME shortcut command, run test copies of the app, and confirm the command is untouched.

▸ Bash K=org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:/org/gnome/s...

Input

$ K=org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/echonote/; B=$(gsettings get $K command); export PATH=~/.nvm/versions/node/v24.21.0/bin:$PATH; npx electron-vite build >/dev/null 2>&1 && npx vitest run --config vitest.e2e.config.ts -t "Workflow (A|E|F)" 2>&1 | grep -E "Tests |×"; A=$(gsettings get $K command); [ "$B" = "$A" ] && echo "GNOME shortcut UNCHANGED by test instances" || echo "CHANGED!"; rm -rf /tmp/echonote-e2e-*

Output

     × Workflow A: reference note via the shortcut command, persisted and searchable after restart 128404ms
⎯⎯⎯⎯⎯⎯⎯ Failed Tests 1 ⎯⎯⎯⎯⎯⎯⎯
      Tests  1 failed | 2 passed | 5 skipped (8)
GNOME shortcut UNCHANGED by test instances

Agent

The fixes are released as v1.0.5: https://github.com/KrAG2000/echonote/releases/tag/v1.0.5. The bundler ran its full pipeline: portability check, tests, all three packages, inspection, checksums, and publishing the release.

What was going wrong

1. The shortcut was slow because my tests had redirected it. Your GNOME shortcut pointed at a test copy's pipe and an AppImage in the project folder, instead of your installed app. Every press therefore took the slow path, about 2–3 seconds before anything appeared, which invites a second press. The log shows exactly that: one press then gives start → stop → start. Fixed: test copies can no longer touch your shortcut. I verified this by running the full end-to-end suite and checking your shortcut was unchanged before and after.
2. The popup disappeared mid-recording, then came back. If an earlier note finished organizing while you were recording, the app hid the popup after 4 seconds without checking whether you were still recording. Fixed: it now stays up for the whole recording.
3. Two old copies were running. Both were version 1.0.2, from before your upgrade, so the 1.0.4 features (silence stop, red dot) weren't active. One had started after you deleted ~/.config/EchoNote, which also removed the lock that normally prevents a second copy. I shut both down cleanly, and both logged a normal stop.

All 102 unit tests and all 8 end-to-end tests pass.

To try it (two commands):
bash
sudo dnf install ./dist/echonote-1.0.5.x86_64.rpm

Then start EchoNote once from Activities. On startup it rewrites your GNOME shortcut to point at the installed app with the fast path, about 20 ms per press.

What I couldn't check: pressing the physical keys myself. After you start 1.0.5, one press of Alt+Shift+Space should start recording almost instantly. Then either press again or just stop talking for 5 seconds. If it still repeats, tell me and I'll read the new log.

It's about 10:45 AM, so ~1h45m to the 12:29 PM deadline. After this check, record the video and publish.

Show more

33 of 33 messages

Prize Categories

Best Use of Gemma

Gemma 4 E2B is the main AI model in EchoNote. It runs locally and turns the things I speak into tasks, reminders, ideas or notes.

I tested it against other models and selected it because it was as accurate as the best alternative (Qwen2.5 1.5B), while still being usable on CPU. Qwen2.5 1.5B was equally strong in my tests. I used Gemma for this project so I could also enter the Gemma prize category. Just being honest