Use the same Mac, microphone, room, and fictional script for each model. Open SuprFlow → Speech → Speech recognition model. The installed 1.0.0 (12) screen shows Small, Small (English), Medium, and Medium (English) cards; the English variants carry an EN badge. A server-provided configuration can change which cards are available, so compare only models actually shown in your signed app. Choosing a larger model from its label alone does not prove it is best for your accent, vocabulary, speed target, or Mac.
The installed Speech screen describes Small at about 460 MB and Medium at about 1.5 GB. Speech → Storage lists downloaded models and labels the loaded one Active. The displayed sizes are not a measured total transfer for your installation. A new model may require another download; the capacity guide covers setup planning.
Build a repeatable test card
Dictate a 60–90 word paragraph three times on each available model. Include a fictional surname, acronym, amount, date, negation, and a short list. For example: “Corwin Capital did not approve the $12,500 transfer on 14 June. Maya will verify the invoice by Thursday. The next three checks are identity, amount, and authorization.” Keep the selected formatter, snippets, Dictionary entries, and destination fixed. If possible, inspect the raw transcript separately from cleaned and formatted text so a rewrite does not mask a recognition mistake.
| Measure | Record for each run |
|---|---|
| Material errors | Wrong name, amount, date, negation, acronym, or action owner |
| Ordinary corrections | Number of word or punctuation edits before approval |
| Latency | Time from releasing the hotkey to usable text, using the same destination |
| Resource impact | Subjective responsiveness and measurable memory/CPU if IT has an approved tool |
| Model setup | Download time, stored space, load time, and whether the model was already cached |
Rotate model order or repeat on another day so familiarity does not always favor the last model. Use the same speaker and microphone. This small test tells you which model fits your task, not a population-wide accuracy percentage. The upstream Whisper model card describes differences across languages and model sizes, but those results are not SuprFlow-specific benchmarks.
Choose according to the actual work
For English-only notes, compare Small with Small (English), then Medium with Medium (English) if the speed and storage budget permits. For non-English or mixed-language work, choose a multilingual model and run the multilingual test. The model screen warns when an English-only model is active with a non-English language. If accuracy in numbers or negations remains poor, keep a human review step and consider microphone, room, and phrasing before assuming a model switch alone will fix it.
Questions people ask
Is Medium always more accurate than Small for my work?
The app labels Medium as high accuracy, but the only meaningful answer for a buyer is a repeated test on the actual language, accent, microphone, and note type. Do not invent an improvement percentage.
Why can't I see every model listed here?
The app filters cards through available-model configuration. Your signed version or account may show a subset. Use the visible cards and record the app version.
Can I test a model without downloading it?
An inactive uncached model needs its assets before it can run locally. Budget storage and network time, and confirm the model reaches Active before scoring its output.
Related guides
- Plan first model download space and bandwidth
- Test multilingual and mixed-language dictation
- Why a Dictionary correction may not change a word
- Run a fair Mac dictation app comparison