BENCHMARKS

Pick a setup, quickly.

Quick nudges on which setup to pick, from synthetic speech, screenshots and dictations. The app shows the same results on its Benchmarks page.

A quick vibe check from 2026-10-19 with Phorminx 0.2.3, on a typical desktop, not a lab test ·6 features · Raw data

Instant (Vosk) for live text, qwen2.5:7b for formatting

Fastest: Instant (Vosk). Live text after about 0.8 s.
Most accurate: A tie. 4 engines got clear English speech about 0% wrong; pick by speed.
Portuguese: Accurate base. Instant is English-only; this got about 2% word errors.
Fastest: qwen2.5:1.5b. About 0.5 s per dictation.
Best quality: Close call. qwen2.5:1.5b, gemma3:4b, qwen2.5:7b scored within 3 points of each other (about 96/100).
Best at lists: gemma3:4b. About 4.1 list lines per dictation in Structured and Custom, against 1.1 for the least list-happy model.
Good default: qwen2.5:7b. Feels instant (1.6 s), scores 95/100 and turns spoken lists into lists.
Second pass: Usually skip it. It changed about 6% of words and took 2.0× as long.
Fastest summary: qwen2.5:1.5b. A one-minute meeting summarized in about 1 s.
Quickest stop: Instant (Vosk). Transcript saved about 0.1 s after Stop.
Screenshots: On-device OCR. A full-HD screenshot reads in about 1.5 s with 0% character errors.
Fastest answers: qwen2.5:1.5b. About 3 s per answer.
Expect: A slow first contact. About 72 s to connect and add a contact over Tor, then about 0.8 s per message.
Startup: Quick. The window is ready about 1.3 s after launch.

Key numbers
Instant: first live text	0.81 s
Accurate base.en: end of speech to text	1.27 s
qwen2.5:1.5b: Balanced, 1 pass	0.51 s
gemma3:4b: Balanced, 1 pass	34.3 s
qwen2.5:7b: Balanced, 1 pass	1.54 s
OCR: full-HD screenshot	1.50 s
Tor: message round trip	0.81 s
App: warm startup	1.26 s

Turn on JavaScript for the charts, filters and every setup.