Meeting notes that
run on your machine.
Speech recognition and speaker attribution run entirely on your computer. The audio never leaves the device — not because we promise it, but because there is no route out.
summo — recording · 12:04
rtf 0.11 · 9× realtime · ram 312 MB · gipformer-65m · offline
Three principles, none of them negotiable
Every engineering decision in Summo follows from these.
Press record and it records
Your data is your files
This is what it looks like
Real screenshots of the running app, not mockups.
Measured, not promised
Every figure below has a command that reproduces it. Where it was measured is in the caption; you can run it yourself.
- 2.4%
- summo-bench asr
- 9×
- summo-bench asr
- 0.94
- summo-bench vad
- 30ms
- summo-bench vault
| Dataset | WER | CER | RTF | Note |
|---|---|---|---|---|
| Fleurs VI (15 clips, 146s) | 2.4% | 1.7% | 0.024 | read speech |
| A real meeting recording (34s) | — | — | 0.107 | includes re-decoding for live text |
| Meetings | On disk | Search (8 threads) | List the vault | Index size |
|---|---|---|---|---|
| 100 | 7 MB | 4 ms | 1 ms | 8 MB |
| 1,000 | 65 MB | 30 ms | 5 ms | 80 MB |
| 5,000 | 327 MB | 140 ms | 26 ms | 402 MB |
Against the cloud services
Not every comparison table is fair, so this one only lists things checkable from their own pricing pages.
| Summo | Otter.ai | Fireflies | |
|---|---|---|---|
| Audio leaves your machine | Never | Yes | Yes |
| Works with no network | Yes | No | No |
| A bot joins the call | Not needed | Yes | Yes |
| Monthly minute cap | None | Yes | Yes |
| Where your data lives | Markdown files | Their servers | Their servers |
| Open source | AGPL-3.0 | No | No |
| Price | Free | $16.99/user | $18/user |
Models, like `ollama pull`
There is no "basic" and "premium" tier. Each model is a standalone manifest stating its size, the RAM it needs, its measured speed per CPU class, the languages it covers and its licence. Summo measures your machine and suggests one — you can pick a different one.
Vietnamese, English, and 97 other languages
An open registry, with nothing locked to us
Pricing
Today there is exactly one thing to pay: nothing. The full application is open source and free, with no limits. The two below are the sync service — not built yet, not for sale, and listed here so you know in advance what would ever cost money and what never will.
Free
$0
forever, open source
DownloadSync
—
end-to-end encrypted, across your devices
Tell me when it existsTeam
—
for organisations
Get in touchWho makes Summo
Techainer — an AI company in Hanoi, founded in 2019.
Techainer builds AI for work that has consequences: video surveillance analytics (Horus), digital identity verification (Master eKYC), document digitisation (DocChain). Over 2 million eKYC registrations and 100,000 banking documents digitised — which is to say our systems run where a mistake is somebody's afternoon.
Summo started as an internal need: meetings held in Vietnamese, minutes written by hand, and nobody willing to upload a recording of an internal meeting to a company in another country. The cheapest way to settle that argument is to make sending it impossible — so recognition runs on the machine, and the source is open so anyone can check that it does.
Thirty seconds to find out
Unpack it and run `./summo serve`. On first launch it measures your machine and asks which model to fetch — nothing is recorded until you press record.
Questions people ask
"The audio never leaves" — how would I check that?
Three ways. One: the source is open, recognition lives in crates/summo-asr, and there is no HTTP client in it. Two: unplug the network and record — everything still works. Three: firewall the app and the transcript is unchanged. Only summarising and answering questions need the network, and the settings screen states whether the endpoint you chose is local or remote.
Does a bot have to join the call?
No. Summo captures your microphone and your system audio directly, so it works with Zoom, Meet, Teams, a phone call, or people in a room — anything your machine can hear.
How accurate is Vietnamese?
2.4% WER on Fleurs VI with the gipformer-65M model. But Fleurs is clear read speech; a real meeting is much harder, and we have not published a number for real meetings because there is no hand transcript to check against yet. When there is, it will sit beside this one.
Does translation run locally?
Yes. As of this version, translation into 46 languages runs on a model inside the app — no Ollama, no network. Only summarising and question-answering call a language model, and you choose that endpoint.
What if my machine is slow?
On first launch Summo measures it for fifteen seconds and picks a pair of models to match. The default Vietnamese model is 73 MB and runs 9× faster than realtime on four CPU cores — no GPU.
Where is my data?
A folder of Markdown in your application data directory. Open it in Obsidian, grep it, put it in git. Audio is stored as Opus and deleted after 30 days by default while the transcript is kept — both configurable.
How can it be AGPL and still be a business?
The app and the core are AGPL-3.0, so nobody can take it and run a closed service on it. What would be paid for later is the sync service, which lives in its own repository and is not needed to run the app. A company that needs to be free of the AGPL buys a commercial licence.