freshfloAccuracy · how we measure
95% exactly right.
Here’s what that means.
We test every release on 355 real-world dictations and publish how. The numbers are ours, so are the tests: read them as a careful self-assessment, not an independent benchmark.
94.9%337 of 355 cases
Freshflo for Mac 0.22, measured October 24, 2026, with the default cloud engine (Groq). A case passes only if every check passes.
| Test set | What it checks | Cases | Passed |
|---|---|---|---|
| Everyday dictation | Fillers, self-corrections, numbers, names, tech terms, questions, lists, emails, AI prompts, long messages, and text that should be left alone | 98 | 93 (95%) |
| Grammar | Agreement, tense, articles, prepositions, homophones and Indian-English phrasing | 76 | 75 (99%) |
| Hindi and Hinglish → English | Faithful translation: who’s speaking, tense and meaning kept; no Hindi left; corrections applied | 50 | 50 (100%) |
| 25 other languages | Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Urdu, Spanish, French, German, Japanese, Chinese, Korean, Russian and more, into English and back | 82 | 80 (98%) |
| Spoken clips | Audio through the full app: speech recognition, then clean-up. Indian-English, Hindi and 16 other voices | 49 | 39 (80%) |
How a case is judged
- Each case lists words that must appear (“Wednesday”, “₹2,50,000”) and words that must not (the filler, the corrected-away “Tuesday”, any leftover Hindi).
- Translated text must really be in the target language, in the right script.
- Several right answers are accepted where English allows them (“late” or “delayed”).
- The text cases start from what speech recognition produces, so they test Freshflo’s editing; the spoken clips test the whole path from audio.
Limits, honestly
- We wrote the tests. They reflect the product we’re building, which is a bias.
- Spoken clips use computer voices, which are harder to understand than most people; real speech usually does better, accents vary.
- Language models aren’t perfectly repeatable: a case can pass on one run and fail on the next.
- Results are for the cloud engine. On-device mode is private and offline, and somewhat less accurate.
- It’s a beta. Tell us where it gets your speech wrong: help@ Freshflo.