Evidence, with boundaries
Benchmarks that say what they measured.
A speech model number is not the same thing as release-to-insert latency. We publish the scope, hardware and exclusions so you can compare results without a marketing shortcut.
Accuracy and speed
- Rambleproof, or the engine family it runs
- Other tools and engines
19% fewer words wrongwith Parakeet v2 than with Whisper large-v3, on the same audio. These are speech engines, not apps.
Same audio, independent leaderboard
* Measured elsewhere, on different audio
Lower is better. Rows marked * or † were measured elsewhere. Parakeet v2 makes about 14 fewer word errors per thousand than Whisper large-v3 on an independent leaderboard. It compares two engines, not two apps. Superwhisper can run Parakeet too, and Rambleproof's default English model is Parakeet Unified, a different configuration. Wispr Flow and Apple come from other sources.
Sources and the numbers as a table
- *Not the same test as the unmarked bars: different audio, in a different year. Wispr Flow 8.7% is published by Superwhisper, a direct competitor, and another rival puts it at 6.2%. Apple is the University of Washington's figure for single words read aloud, measured in 2022: 14% for the best-served speaker group, 24% for the other three.
Sources: Open ASR Leaderboard, English short-form. * Scott, Liang, Wassink and Levow, University of Washington (arXiv:2510.22495).
The leaderboard measures speech engines, not whole apps, averaged over 8 datasets: AMI, Earnings22, GigaSpeech, LibriSpeech clean, LibriSpeech other, SPGISpeech, TED-LIUM v3, VoxPopuli. Before 1.0.9, Rambleproof's own end-to-end time was 1.32s median over 37 dictations (key release to cleaned text at the cursor, MacBook Air M4, macOS 26.5.1); 1.0.9 streams transcription while you talk and brings that to about 1 second.
| Tool | What was measured | Accuracy | Measured by |
|---|---|---|---|
| Parakeet v2 | Parakeet TDT 0.6B v2 | 6.05% | Open ASR Leaderboard, English short-form |
| Whisper large-v3 | Whisper large-v3 | 7.44% | Open ASR Leaderboard, English short-form |
| Wispr Flow | engine not disclosed | 8.7% | Superwhisper, a direct competitor |
| macOS Dictation | Apple Speech API, single words, 2022 | 14.0% | University of Washington |
What everyone else publishes, and who published it
The full set of figures the dictation industry publishes about itself, including the ones the chart above draws. Every one was produced by a company selling one of the products being scored, and none publishes a corpus you could use to check the result, which is why they are charted apart from the independent benchmark rather than beside it. Where two of them measure the same product, both are listed, disagreement included.
| Tool | Figure | Value | Published by |
|---|---|---|---|
| Wispr Flow | Word-error rate | 6.2% | LumeVoice A competing dictation app, ranking itself first in the same table. No corpus or method published. |
| Superwhisper | Word-error rate | 3.6% | LumeVoice Same table, same publisher. Disagrees with the independent leaderboard figures for the Whisper and Parakeet engines charted above, both engine families Superwhisper offers. |
| macOS Dictation | Word-error rate | 20.9% | LumeVoice Same table. Other dictation blogs put the same product between 7.8% and 18%, which is the spread you get when nobody shares a corpus. |
| Wispr Flow | Word-error rate | 8.7% | Superwhisper A direct competitor measuring Wispr Flow. Publishes more of its method than most, but is still a contestant scoring a rival. |
| Wispr Flow | Time to final text | 0.80 s | Embertype A competing dictation app. No trial count and no method stated. |
| Superwhisper | Time to final text | 1.40 s | Embertype Same post, same absence of method. |
| macOS Dictation | Time to final text | 0.30 s | Embertype Same post. Yapper measures the same product at 760 ms, which is 2.5 times slower. |
| macOS Dictation | Time to first character | 760 ms | Yapper A competing dictation app, but the best method of the set: 100 trials per app on an M3 Pro, timed off 120 fps screen capture. It measures the first character, not the finished text. |
The only numbers on this page we measured ourselves are Rambleproof's, and we say exactly how: in 1.0.9, about 0.3 seconds from letting go to raw text and about 1 second to cleaned text at the cursor, on a MacBook Air M4 (1.32 seconds median across 37 dictations before 1.0.9). The one competitor timed the same way is Wispr Flow, from its own logs on the same Mac, shown in the chart's speed tab.
Current measured result
| Metric | 150 ms median local ASR inference |
|---|---|
| Corpus | 188 English clips |
| Hardware | M4 MacBook Air |
| Included | Model transcription inference for the prepared clip |
| Excluded | Recording, audio-tail capture, cleanup, target validation, accessibility insertion and clipboard fallback |
Local ASR inference only. It excludes recording, audio-tail capture, cleanup, accessibility insertion, and clipboard fallback.
Reliability suites
The automated corpus includes 120 silence cases and 24 cleanup safety dimensions. The application also exercises ordered commits, retention, interrupted recordings, shortcut handling, cleanup fallback and insertion policy. Counts can grow; behavior and failure policy matter more than a frozen test total.
From letting go to text on screen
Release to text in Rambleproof 1.0.9: about 0.3 seconds to raw text and about 1 second to cleaned text at the cursor, measured on a MacBook Air M4 (1.32 seconds median across 37 dictations before 1.0.9). That covers the whole trip: the end of the recording, transcription, cleanup and pasting at the cursor. It is one Mac and one person, so treat it as a typical result, not a promise for every setup. To see how any dictation app handles your own voice and work, run the six short tests in our guide to testing a Mac dictation app.
Acceptance rule
A result cannot be promoted to the landing page unless the artifact, corpus revision, machine, percentile, sample count and measurement boundaries are recorded together. Failed cleanup is a fallback result, not a silent pass. Failed insertion must leave recoverable text on the clipboard and visibly tell the user.