Stream-1 Mini vs S1-mini
On our dictation benchmark, Stream-1 Mini scored 89%, while S1-mini scored 83%.
Updated
Superwhisper offers its S1-mini model for local cleanup. As far as we know, it's the only other Windows dictation app with a fine-tuned cleanup model. Both models turn raw speech-to-text into clean writing on your own computer. We ran both on the same 1,930 dictation tests to be transparent about how our model stacks up against the competition.
Here's what we found.
Dictation cleanup benchmark
1,930 tests
| Benchmark |  Stream-1 Mini Whisperstream |  S1-mini Superwhisper |
|---|---|---|
| OverallOverall, share of all tests | ||
Tests passed No error of any kind | 89.3% | 83.3% |
Critical errors Changed what you meant | 4.9% | 6.8% |
Major errors Cleanup or formatting failed | 5.6% | 9.5% |
Minor errors Small defects | 0.2% | 0.4% |
| By kind of dictationBy kind of dictation, tests passed | ||
Self-corrections Uses the final choice after “no” or “I mean”. 624 tests | 88.9% | 84.8% |
Not a correction Leaves “no” alone when it is part of what you meant. 625 tests | 91.8% | 91.2% |
Everyday cleanup Mixed everyday dictation, cleaned without losing anything. 235 tests | 89.8% | 80.9% |
Bulleted lists Saying “bullet point” before each item. 52 tests | 94.2% | 0.0% |
Numbered lists Steps like “first” or “step one” as a numbered list. 93 tests | 95.7% | 88.2% |
New line and paragraph Saying “new line” or “new paragraph”. 50 tests | 100.0% | 90.0% |
Email layout Greeting and sign-off on their own lines. 509 tests | 87.8% | 85.1% |
Dictated requests “Summarize this” or a question gets typed out, not answered. 19 tests (small sample) | 89.5% | 89.5% |
Both models got the same tests. Bold marks the better result in each row.
Example outputs
These are unedited outputs from each model.
Stream-1 MiniPassGood morning,
The grant application is due Tuesday, and we still need two reference letters.
Book the conference room for Thursday. Actually anyway no, make that Friday.
S1-miniCriticalGood morning.
The grant application is due Tuesday Friday, and we still need two reference letters. Book the conference room for Thursday. Actually anyway no, make that Friday.
Moved the grant deadline from Tuesday to Friday. The correction was only about the room booking.
How we tested
- Critical
- The cleanup changed or lost what the speaker meant.
- Major
- The meaning survived, but the cleanup or formatting was wrong.
- Minor
- Small defects that a reader would probably let slide.
- The 1,930 tests come from our development and regression suite.
- S1-mini here is the public v076 checkpoint.
Start speaking.Completely free.Fully local.
Open source, so you can read every line of code.