Dokumentacja
Whisper, Parakeet or Nemotron: which speech model for dictation
Almost every local dictation program runs one of three model families. They differ more than their names suggest: in how much memory they take, how long a sentence takes on a processor, whether they punctuate, and what languages they cover. HoldToType ships all three and a few relatives, so the numbers below come from the same program on the same machine.
Zaktualizowano · Vitalii Yemets
The three families in one paragraph each
- Whisper (OpenAI, run through whisper.cpp) is the generalist: 99 languages, translation into English by design, sizes from 74 MB to over a gigabyte. It hears the whole phrase after you let go and answers slowest of the three. Punctuation is partial.
- Parakeet TDT 0.6B v3 (NVIDIA, run through sherpa-onnx) is the sprinter: 25 European languages, punctuation built in, a sentence in well under a second on a CPU, 670 MB on disk. It does not translate.
- Nemotron 3.5 (NVIDIA, streaming) is the one that types while you talk: partial text every few hundred milliseconds, 40 languages, punctuation built in, 651 MB. It needs a little more headroom than Parakeet.
Size and memory
Disk size is what you download once; memory is what the program holds while a model is loaded. HoldToType estimates memory as 1.2 times the file for the sherpa-onnx models and 1.5 times plus 60 MB for Whisper, and shows the estimate next to each model.
- Whisper Tiny 74 MB on disk, Base 142 MB, Small 466 MB, Medium q5 539 MB, Large v3 Turbo q5 574 MB.
- Parakeet TDT 0.6B v3 670 MB; Nemotron 3.5 651 MB; Canary 180M 198 MB; GigaAM v3 232 MB; Qwen3 ASR 0.6B 937 MB.
- In memory: Whisper Medium about 870 MB, Parakeet about 800 MB, GigaAM about 280 MB.
Speed on a processor
The one measurement HoldToType publishes with numbers is the Russian case, because it is the starkest: GigaAM v3, a narrow sherpa-onnx model, turned an eleven-second phrase into text in 0.47 seconds where Whisper Medium q5 took 11.6 seconds, using 277 MB of memory instead of 814. Parakeet behaves like GigaAM: a sentence is text well under a second after you let go. Whisper on the same hardware takes roughly the length of the phrase. Nemotron does not wait at all: the words arrive while you speak, and the final text is ready almost as soon as you release the keys.
The machine was an ordinary laptop processor with no graphics card; all three families run without one, which is the point.
Punctuation, languages, translation
- Punctuation: Parakeet, Nemotron, GigaAM v3 and Canary punctuate by themselves. Whisper does it partly; HoldToType's post-processing finishes the job with a small model on your disk.
- Languages: Whisper 99. Nemotron 40. Parakeet 25 European ones, including Ukrainian, Polish and Russian. GigaAM Russian only. Canary English, German, Spanish and French.
- Translation: Whisper translates into English by design and into the other interface languages when told the speech is already in the target one; Canary translates between its four languages. Parakeet and Nemotron transcribe only, so HoldToType keeps the best installed Whisper standing by and the plate offers to hand a phrase to it.
- Live text: Nemotron only.
Which one to pick
HoldToType keeps one model per language and switches when you change the language, so you do not have to pick one for everything. The models page has the catalogue and a paragraph per language; the languages page shows where each language's model is set.
- Everyday dictation in a European language: Parakeet. Fast, punctuated, accurate.
- You want to see the words as you speak: Nemotron 3.5.
- Russian: GigaAM v3, and nothing comes close on speed.
- Translation, or a language outside the 25: Whisper Medium q5, or Small on a weak machine.
- An old laptop: Canary 180M for English, German, Spanish and French; Whisper Base elsewhere.
- Accuracy above everything and memory to spare: Qwen3 ASR.
Trying them yourself
- Install from the download section and let the wizard fetch one model.
- Open Languages & models, add a second model for the same language, and switch between them; the log in the System tab prints how many seconds each dictation took.
- Say the same sentence to each. The difference between Whisper and a narrow model is visible without a stopwatch.
Questions
Do I need a graphics card?
No. HoldToType runs every model on the processor. A GPU would make Whisper faster in other programs; the narrow models are already fast without one.
Which is the most accurate?
For clean speech in a major language the three are close. Qwen3 ASR is the most accurate in the catalogue and the heaviest. For names and jargon, the dictionary matters more than the model.
Can I bring my own model?
Yes. Drop a Whisper GGML file or a sherpa-onnx model folder into the models folder and it appears in the list.
Try all three on your own machine
One install, nine models, and the log tells you the seconds.