Parakeet v3 against Whisper, language by language

The figures

"Words wrong" is the word error rate: every word the engine swapped for another, left out or added, counted against the number of words in the sentence that was actually read, across all 80 clips. So 7.4% in German means a little over seven words in every hundred came out wrong.

The gap is widest where Whisper small struggled most. In Dutch and French it got about one word in five wrong, and Parakeet v3 roughly halves that. In Italian Parakeet gets about a third as many words wrong. In German, Spanish and Portuguese the improvement is smaller but still there: about a third fewer mistakes in German, and about two in five fewer in Spanish and Portuguese.

The time column is the median time Parakeet v3 took to decode one clip on that machine: half the clips were quicker, half were slower. It is the decode alone. It does not include recording, the moment between letting go of the key and the engine starting, or typing the text into your app.

The Spanish figures are for Latin American Spanish, because that is FLEURS's Spanish split, and the Portuguese figures are for Brazilian Portuguese for the same reason. Neither says anything about Spain's Spanish or European Portuguese.

English stays on Whisper

Parakeet v3 is twice as fast in English, and that was tempting. But it gets more words wrong, and it loses badly on the thing English users of Deixis lean on most, which is the dictionary. Whisper reads your list of words before it starts decoding, so a product name, a colleague's surname or a library you have taught it is likely to come out spelled your way. Parakeet hears a dictionary term only by its sound. In a check on 34 names from the Deixis owner's own dictionary, Whisper spelled 20 of them right and Parakeet 10. That check is small and is not a benchmark, but the direction is clear enough that English stays where it was.

The same weakness applies to the six other languages: Parakeet will get unusual names right less often than Whisper does in English. We chose it anyway, because the alternative in those languages was a model that got far more of the ordinary words wrong.

How it was measured

• The audio. Google's FLEURS dataset, test split, which is sentences read aloud. Each language uses the dataset's own regional split, named in the table: de_de, nl_nl, fr_fr, it_it, es_419 and pt_br.
• The clips. 80 per language: the 20 shortest clips in the split, plus 60 drawn at random with a fixed seed of 7, so the same 80 can be drawn again.
• The scoring. Words wrong, counted after lower-casing both the engine's text and the reference and dropping punctuation. A missing comma or capital letter is not counted; a wrong word is.
• The machine. The processor of one 12-core desktop PC, an AMD Ryzen 7900X, running the engines on 11 threads, on 3 and 4 November 2026. Not a graphics card, and not a laptop.
• Parakeet v3. Run through sherpa-onnx 2.0.1, then run again on 4 November through the Deixis app's own core, the code that does the work when you dictate, which agreed with these figures to within 0.4 points.
• Whisper. The multilingual small model, measured again on 4 November after two fixes to the way Deixis decodes with it, so the comparison is with Whisper as it runs now rather than an earlier, weaker run. English uses small.en, the English-only model.

The one-in-ten rule

Deixis offers a language for speech to text only when an engine gets fewer than one word in ten wrong on this test. Past that point you spend more time correcting than you saved by speaking. The six languages above clear it with Parakeet v3, and English clears it with Whisper.

We measured three more and do not offer them. The best figure we measured for Afrikaans was 36% of words wrong, for Zulu 46%, and for Xhosa 42%. That is somewhere between one word in three and one word in two, which is not something we would ask anyone to dictate with.

What was not measured

These are gaps in the evidence, and the pages that quote these figures say so too.

• Regional accents. Swiss German, Flemish, Quebec and Belgian French, Spain's Spanish and European Portuguese were not measured. The figures are for the FLEURS split named beside each language and nothing wider.
• Free speech. FLEURS is read aloud from written sentences. The way people dictate an email, with hesitations, restarts and half-sentences, is harder, and it was not measured.
• Noisy rooms. Nothing here tested background noise, other voices or a poor microphone.
• Mixed languages. Deixis dictates in one language at a time, the one you choose. Switching language inside a sentence was not measured, and no page claims it.
• Laptops and the Mac. Every time on this page comes from one desktop PC running Windows. A laptop is slower, and nothing here was measured on a Mac.
• End-to-end latency. Only the decode was timed. How long it takes from letting go of the key to seeing text in your app was not measured on this page.
• Everything after recognition. Cleanup, the dictionary, the vocabulary packs and the punctuation Deixis adjusts to the field you type into are tuned for English, and nothing here measures them in the other six languages.