Blog
We measured Wispr Flow in five of its 100+ languages. English scored 95.5%, Hungarian 76.7%.
7 min read
Emil Lykke Grann
Wispr Flow's homepage says "100+ Languages", and a post on Wispr's own blog says "95%+ accuracy". From what it sounds like they are claiming an average of 95%+ accuracy across those 100+ languages. To see if this was actually the case I decided to benchmark them on 2,000 clips of audio.
Clean English came back at 95.52% word accuracy and Hungarian at 76.70%. Both languages sit under the same "100+ Languages" headline.
While they do score 76.70% on Hungarian, they perform very well on Danish, scoring 94.55% against 88.92% for our best Danish model.
Where Wispr's two claims actually come from
The 95% figure appears twice in a post on Wispr's blog from 16 December 2025, and both times it is attached to platforms rather than languages:
"Wispr Flow leads with AI-powered voice typing that works across every platform at 95%+ accuracy."
That post is a listicle written by Wispr that ranks Wispr first against competitors, with no methodology and no facts about how they measured their own accuracy. The same page quotes a competitor at "3.2% WER on Librispeech clean", which is exactly what a checkable claim looks like.
The 100+ Languages claim lives on Wispr's homepage as well as multiple blog posts.
What a checkable accuracy number needs
Word error rate counts every wrong, missing and invented word against the number of words actually spoken, and word accuracy is simply 100% minus that rate. Two details decide whether such a number means anything.
The first is how you add it up. We total the word errors across the whole run and divide by the total number of reference words once, rather than averaging per-clip percentages, which would let a three-word clip count as much as a forty-word one.
What we measured
Wispr Flow 1.6.774 ran on an Apple M4 Max with 36 GB of RAM, on macOS 26.6.2, with Context Awareness off and an empty dictionary. Each of the five datasets got 403 clips, the first three as unscored warmups, which leaves 400 scored clips per dataset and 2,000 in total.
| Dataset | Language | Wispr Flow word accuracy |
|---|---|---|
| LibriSpeech test-clean | English | 95.52% |
| LibriSpeech test-other | English, noisier | 93.50% |
| FLEURS es_419 | Spanish | 96.05% |
| FLEURS da_dk | Danish | 94.55% |
| FLEURS hu_hu | Hungarian | 76.70% |
| Pooled | 91.76% |
Pooled, that is 3,453 word errors against 41,912 reference words.
Wispr Flow's 95%+ figure roughly holds on clean English which scored 95.52% and Spanish 96.05%, and the harder English set, landed at 93.50%.
Outside the list things stopped being predictable, and not in the direction I expected. Danish is off the list and scored 94.55%, a shade below clean English. Hungarian scored 76.70%.
Flow wins Danish, and it does not win narrowly. Hviske v5 Tiny is a Danish-only model, and the last thing we published about it was that it beats every multilingual OpenAI Whisper model we have measured on Danish. That is still true. It also loses to Wispr Flow by 5.63 points on exactly the same audio. Hviske v5 Tiny scored 88.92% against Wispr's 94.55%.
Speed: Flow is fastest on English and slowest on Danish and Hungarian
Anything our setup still does inside the timed window is measured rather than guessed, and it comes to a median of 8.6 ms per clip. The metric is milliseconds of waiting for every second of audio you dictate, pooled over the clips that answered.
| Response ms per audio second | English | English noisy | Spanish | Danish | Hungarian | Pooled |
|---|---|---|---|---|---|---|
| Wispr Flow 1.6.774 | 73.3 | 74.5 | 50.3 | 213.3 | 212.0 | 132.4 |
| Whisper Large V3 Q5_0 | 117.9 | 125.4 | 78.8 | 90.6 | 90.0 | 96.4 |
| Whisper Large V3 Turbo Q5_0 | 77.5 | 82.9 | 47.0 | 53.4 | 51.5 | 59.0 |
| Parakeet TDT 0.6B v3 | 25.1 | 25.6 | 15.4 | 16.7 | 15.9 | 18.6 |
Flow is the fastest thing in the table on both English sets, Whisper Large V3 Turbo Q5_0 is faster on Spanish, and Parakeet TDT 0.6B v3 is roughly three to thirteen times faster than Flow everywhere.
How you benchmark a product that has no API
Wispr Flow has an API, but not available to me. It listens to a microphone and pastes text into whatever window is focused, so every part of our setup works around one of those facts.
corpus WAV
│
▼
Swift bridge ──▶ CoreAudio default output ──▶ BlackHole 2ch (loopback)
│ │
│ Option+Z via CGEvent ▼
├──────────────────────────────▶ Wispr Flow (mic = BlackHole)
│ │
▼ ▼
focused NSTextView ◀─────────────────── Flow pastes text
│
▼
whisper-basic normalisation ──▶ Levenshtein alignment ──▶ WER
The only supported way in is a microphone, so we use a virtual one. BlackHole 2ch is a loopback audio driver. Set it as the system's default output and anything you play goes straight to its input side, which is what Flow listens to. There are no speakers and no room anywhere in the path.
Flow only accepts audio input via virtual microphone, so we used BlackHole for loopback and triggered transcription with a simulated Option+Z keypress. Transcribed text was pasted into a focused text box, and was considered final when unchanged for 750 ms. Scoring used the same normalisation and WER process as Codictate.
The updated benchmark system ensures each run tracks unique audio files, fixing double-counting from repeated sentences in FLEURS.
The Flow run alone took almost 8 hours for 2,000 clips, most of which was real-time audio. The whole batch of 30,000 clips (Flow, 13 multilingual, 5 Danish models) took 15 hours, so I had to leave it running overnight. Details and code: github.com/EmilLykke/dictation-benchmark.
Comparison table
So it turns out, according to our sub-optimal benchmark setup on Wispr Flow, the only language it outperforms Codictate in is Danish.
| Model | English | English noisy | Spanish | Danish | Hungarian | Pooled |
|---|---|---|---|---|---|---|
| Wispr Flow 1.6.774 | 95.52% | 93.50% | 96.05% | 94.55% | 76.70% | 91.76% |
| Whisper Large V3 Q5_0 | 96.78% | 95.18% | 97.20% | 87.12% | 85.20% | 92.60% |
| Whisper Large V3 | 96.69% | 95.28% | 97.17% | 87.38% | 85.16% | 92.64% |
| Whisper Large V3 Turbo | 96.59% | 95.02% | 96.64% | 85.68% | 83.14% | 91.74% |
| Parakeet TDT 0.6B v3 | 96.08% | 94.39% | 94.31% | 80.61% | 81.33% | 89.60% |
| Hviske v5 Tiny Q5_0 | 88.92% |
What this benchmark does not tell you
We measured five languages, not a hundred, so we have nothing to say about the other 95. A virtual audio device is also not a room. Loopback audio is cleaner than a person talking into a laptop with a fan running, and that removes a category of difficulty from both products equally, which is fine for ranking them and wrong for predicting your own experience. LibriSpeech and FLEURS are read speech rather than someone dictating a Slack message and changing their mind halfway through, and read speech flatters everything.
We also measured the base product rather than a personalised one. Wispr says accuracy improves with use, and a Flow that has learned your jargon is not what we tested. The speed columns are not the same measurement, for the reasons above, so rank within a column freely and treat gaps under a factor of two between the products as noise. And this is one machine, one run and a 45 second answer limit of our own choosing, with no repeats and no confidence intervals.
FAQ
How accurate is Wispr Flow? We measured 91.76% pooled word accuracy across five languages and 2,000 clips, or 91.89% if you leave out the two clips that returned nothing. By language that is 95.52% on LibriSpeech test-clean, 93.50% on test-other, 96.05% on FLEURS Spanish, 94.55% on Danish and 76.70% on Hungarian. It is read speech with personalisation off, so treat it as a ranking rather than the number you will see yourself.
Does Wispr Flow's 95%+ accuracy claim hold up? No. According to our benchmark, it does not. But it does go above in English and Spanish. It just underperforms on Danish (slightly) and Hungarian (significantly).
Is Wispr Flow faster than a local model? On English it is faster than our most accurate model and slower than our fastest, at 73.3 ms per second of audio against 117.9 for Whisper Large V3 Q5_0 and 25.1 for Parakeet TDT 0.6B v3. On Danish and Hungarian it is roughly 213, which is slower than every Codictate model we measured. Response times are not measured the same way for both products: Codictate is timed at the direct adapter call boundary, Wispr Flow is timed from the UI-observed paste which should be kept in mind when comparing the two.
Reproduce it or argue with it
Everything that produced the Flow column is open source at github.com/EmilLykke/dictation-benchmark. You will need a Mac, BlackHole 2ch, a Wispr Flow account and a day of not using your computer. Every Codictate run behind the tables above is on the benchmarks page, and the Danish models have their own write-up in the Hviske v5 Tiny post.
If you speak one of the other 95 languages, the honest answer to "how good is it for me" is that nobody has published the number, including us. The most useful thing either of us can do about that is print the denominator next to the percentage.