How we measured
Same passage. Same Mac. Same stopwatch.
We read one passage aloud 13 times, speaking into PUSHTYPE and Wispr Flow at once, and timed how long each took to hand back the text after the key was released. These are medians of all 13 readings, not best runs. Measured 30 September 2026 on a Mac mini (M4, 24 GB), macOS 27.2, PUSH 8.2.4 (the app is now called PUSHTYPE).
The passage
“Let's move the Q3 planning review to Tuesday, March 3rd at 4:30 PM, since that's like the only slot that works for the Zürich office. Um, we've spent 37% of the $5 million budget, and the new build loads in 15.2 seconds, I mean 1.52 seconds. Everyone who tried it said, quote, it feels like a real improvement, end quote. New paragraph. Could you like send the final numbers before Friday?”
Every phrase tests something dictation apps handle differently: money, decimals, a correction mid-sentence, a filler word, a spoken quotation, a spoken paragraph break, an accented place name and a question. Two casual “like”s test PUSHTYPE's California mode, and “feels like” has to survive it. Each reading took 23–29 seconds.
Speed: release to text
| App | Where it runs | Median | Range |
|---|---|---|---|
| PUSHTYPE · Parakeet Streaming default in English | On your Mac, English only | 0.02 s | 0.019–0.041 s |
| PUSHTYPE · Parakeet Ultra default elsewhere | On your Mac, English and European languages | 0.04 s | 0.041–0.046 s |
| Wispr Flow | Their servers | 0.74 s + 0.13 s network | 0.70–0.98 s total |
Parakeet Streaming, faster than Wispr Flow once their network round trip is counted (32× against their processing alone).
Parakeet Ultra, the same comparison (17× against processing alone). The smaller of the two is the number on our home page.
Both apps are timed from the moment the key comes up. PUSHTYPE transcribes while you talk: Streaming works through the audio continuously, and Ultra finishes each sentence as soon as you've said it, so only the last moment is left when you let go. Decoding a whole 25-second recording from scratch would take Ultra 0.23 s. Wispr Flow's times are its own, read from its local records.
Accuracy: what each one typed
| The passage asks for | PUSHTYPE Streaming | PUSHTYPE Ultra | Wispr Flow |
|---|---|---|---|
| “five million dollars” → $5 million | 13/13 | 13/13 | 12/12 |
| “15.2 seconds, I mean 1.52 seconds” → 1.52 seconds | 13/13 | 13/13 | 12/12 |
| “quote … end quote” → quotation marks | 13/13 | 13/13 | 12/12 |
| “New paragraph” → a paragraph break | 13/13 | 13/13 | 12/12 |
| Casual “like” removed, “feels like” kept | 13/13 | 13/13 | 12/12 |
| “Um” removed | 13/13 | 13/13 | 12/12 |
| The question mark | 13/13 | 13/13 | 12/12 |
| 37% and Q3 | 13/13 | 12/13 | 12/12 |
| March 3rd | 13/13 | 13/13 | 2/12 writes “March 3” |
| Zürich | 0/13 | 0/13 writes “Zurich” | 5/12 |
Counted over every reading: 13 for each PUSHTYPE model, 12 for Wispr Flow (one reading was made without it). PUSHTYPE ran with its default settings plus California mode. The one 37% Ultra missed came from a reading where the reader stumbled on the number.
Where Wispr Flow did better, we say so. It sometimes restores the umlaut in Zürich, and its server rewrites words using the rest of the sentence: once it turned a misheard “bill” back into “build”, and once it put back a word the reader had skipped. PUSHTYPE types what it heard. You can switch off PUSHTYPE's self-correction in Settings ▸ Dictation, because it's the one step that deletes words you actually said.
Type: how fast a suggestion appears
A script typed fresh sentences into TextEdit at 10 keys a second and watched the screen just right of the cursor, so the clock stops when a suggestion is actually visible, not when the model finishes.
Median time from a key to the suggestion on screen.
Keys after which a suggestion was already showing 80 ms later, mid-typing.
From Tab at the end of a sentence to the next suggestion: it's made in the background while you read the current one.
Network connections while suggesting. The model runs on your Mac; we checked that it opens no sockets.
Measured 11 October 2026 on a Mac mini (M4, 24 GB) with the Best model, Gemma 4 E2B. Smaller models are quicker per key; the Pro model is slower.
And how often it's right
We replay 70 everyday openings we wrote (emails, chats, notes) through the same code the app uses, and check whether the first word suggested is the word the writer actually wrote next. With the Best model it is right about half the time straight after a word (26 of 54), and 86% of the time once you've typed two letters of the word (42 of 49). That's before it has learned anything from your own writing.