Compare

Summo against Granola, Humla, Otter.ai and Fireflies — including the rows we do not win. With the measurements: 2.4% Vietnamese WER, 9× faster than realtime, and the command to re-run each one yourself.

Against the apps of the same kind

Not every comparison table is fair, so this one lists only what is checkable from their own pages — including the rows we do not win.

	Summo	Granola	Humla	Otter.ai	Fireflies
Audio leaves your machine	Never	Yes — sent to Deepgram/Assembly	No, with a local model	Yes	Yes
Works with no network	Yes	No	Yes	No	No
A bot joins the call	Not needed	Not needed	Not needed	Yes	Yes
Monthly minute cap	None	Yes, on the free tier	None	Yes	Yes
Where your data lives	Markdown files	Their US AWS VPC	SQLite on your Mac	Their servers	Their servers
Open source	AGPL-3.0	No	MIT	No	No
Runs on Linux	Yes	No	No	No	No
Price	Free	$14/user	Free · $5 for cloud	$16.99/user	$18/user

The two middle columns are apps of the same kind as Summo rather than cloud services — added because they are what you are actually choosing between, and because there are rows we do not win. Granola needs no bot either, and Humla is open source too. Sources: Granola's pricing and security pages ("best-in-class transcription providers (like Deepgram and Assembly)", "Notes are stored in our US-hosted AWS Virtual Private Cloud"), Humla's README and LICENSE. Otter and Fireflies prices are their lowest paid tier at the time of writing, billed monthly.

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

A Vietnamese-specific transducer for the best speed and accuracy; Whisper for everything else and for sentences that mix Vietnamese and English. Switching models mid-recording does not lose the transcript.

An open registry, with nothing locked to us

Manifests are static JSON in a public repository. If Summo's CDN disappears the app falls back to GitHub and then to the original on HuggingFace. You can host the registry yourself.

Techainer/summo-registry

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.

summo — recording · 12:04

rtf 0.11 · 9× realtime · ram 312 MB · gipformer-65m · offline

2.4%: Vietnamese WER (Fleurs VI); summo-bench asr
9×: faster than realtime; summo-bench asr
0.94: voice-activity F1; summo-bench vad
30ms: searching 1,000 meetings; summo-bench vault

Recognition — gipformer-65M, 4 threads

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

That 0.107 includes re-decoding the sentence in progress about ten times so the text follows the speaker. Decoding once costs 0.012 — nine tenths of the budget buys words appearing as they are said, and there is still 9× headroom.

Storage — is a database needed?

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

The answer is no. Five thousand meetings is 327 MB of Markdown, searched in 140 ms — inside the 200 ms where a person starts to notice waiting.

No praise yet, only evidence

Summo has just been opened up and has no users to quote. Inventing a few kind words is easy, and it is the one thing a privacy tool must not do: you cannot open with a person who does not exist and then tell somebody you do not lie about where their words go. So here is what can be checked instead.

“BẠN SẼ DỄ DÀNG HỌC TIẾNG BỒ ĐÀO NHA”

What the end-to-end test read back out of the recording it had just made — verbatim, uncorrected.

A Vietnamese sentence, through a real microphone

Not a sample file: the test machine speaks out loud, records itself, and reads it back. whisper-tiny, on a CI runner, with no network.

pnpm -C apps/web e2e:full-flow

1,340 Rust tests and 384 interface tests

Plus 22 suites driven in a real browser against a real daemon and a real vault, on every push. None of them merely check that the code compiles.

cargo test --workspace && pnpm -C apps/web e2e

The packaged app is started before anybody is offered it

The AppImage is launched under Xvfb in CI: the daemon comes up, the interface is served, and the process is still alive five seconds later. The Android build is installed on an emulator and opened.

scripts/smoke-desktop.sh · scripts/android-smoke.sh