Benchmarks
The numbers, with the conditions.
Most capture apps publish marketing numbers. This page publishes measurements — ours with the machine and date they were taken on, competitors' as attributed third‑party scans — and names what we haven't measured yet.
Method
Measured, or labelled as a target.
Every Snapbak figure below comes from the engineering notes: the bundle measured with du, the dictation bench run on one machine with its conditions stated, and the model comparison on a fixed clip set. Where a number is a design target rather than a measurement, the table says so. Competitor sizes are published scans of the installed apps, not vendor downloads.
- Dictation bench: M4, 16 GB, macOS 15.5, models loaded.
- Accuracy: 60 LibriSpeech test‑clean clips, both engines.
- Bundle size: du -sk on the installed app, October 2026.
- Competitor sizes: third‑party scans (Desktop Insights).
- Nothing here is a target — targets are labelled as targets.
Dictation
Two engines, one bench.
We benchmarked the two open engines a Mac dictation app would realistically ship, on the same clips. Accuracy tied; speed and memory didn't — which is why Snapbak ships Parakeet, and English‑only. The full comparison has its own post.
| Measure (same machine, same clips) | Whisper large‑v3‑turbo | Parakeet TDT 0.6B v2 |
|---|---|---|
| Word error rate, 60 LibriSpeech clips | 2.3% | 2.30% |
| 3 s of speech to text | 0.68 s | 0.045 s |
| 10 s of speech to text | 0.97 s | 0.058 s |
| 30 s of speech to text | 2.51 s | 0.166 s |
| 10 s with a 200‑term dictionary | — | 0.082 s |
| Helper peak memory | 852 MB | 629 MB |
- End to end — key‑up to text inserted in the app you were in: median 0.44 s, worst case 0.53 s, on the bench machine.
- After the helper quits, Snapbak idles at a measured 25 MB. The model weights are the memory; they leave with it.
- The models download on demand — about a 540 MB first‑run fetch — so the installer never bundles them.
Footprint
What it weighs, installed.
Our bundle is a measurement; the field is the published installed sizes from third‑party scans. Different shapes of app — a screenshot utility does less than a suite — which is exactly why the sizes are worth seeing together.
| App | Installed (Mac) |
|---|---|
| Snapbak — measured bundle | 19.8 MB |
| Shottr | 6 MB |
| Handy | 43 MB |
| CleanShot X | 72 MB |
| VoiceInk | 125 MB |
| Superwhisper | 256 MB |
| ScreenFlow | 264 MB |
| Cap | 421 MB |
| OBS | 656 MB |
| Wispr Flow | 697 MB |
| Camtasia | 930 MB |
| Loom | 1,009 MB |
| Screen Studio | 1,299 MB |
Snapbak's installer is targeted at ~15–20 MB — a target, not a measurement; the 19.8 MB above is the measured bundle. Competitor sizes are the published third‑party scans (Desktop Insights, August–October 2026) and change with releases.
Honest gaps
What we haven't measured.
A benchmarks page that skips its own gaps is a landing page. These are the numbers this page is missing, and why.
Why do your competitor sizes differ from vendor pages?
Vendor pages publish download sizes, which compress; the table uses installed sizes from third‑party scans of the actual apps (Desktop Insights, August–October 2026). An Electron app that downloads at 90 MB can install at 1,009 MB — installed is the number your disk lives with.
Why is there no cold‑start number?
Because we haven't measured it to a standard we'd publish: the bench conditions aren't pinned yet, and a number without its conditions is marketing. It will appear here when it's measured — the same standard every number on this page holds to.
What do the dictation benchmarks actually measure?
Two engines, the same 60 LibriSpeech test‑clean clips, one machine (M4, 16 GB, macOS 15.5), models loaded. Word error rate measures accuracy; the seconds measure how long each engine takes to turn a finished utterance into text. End‑to‑end — releasing the key to text inserted in the app you were in — we measure a median of 0.44 seconds on that machine.
Is 19.8 MB the download size?
No — it's the measured app bundle on disk (du -sk, October 2026). The speech models download on demand, about a 540 MB first‑run fetch, so the installer stays small and your disk only pays for the model when you actually dictate.
Snapbak
Check the numbers, then check the app.
Every figure here is reproducible from the stated conditions. Download Snapbak free for the first 100 users and measure the parts you care about yourself — the dictation history even reports its own timings.