Choose a speech model
Choose a shipped Voicetypr speech model by platform, language, download size, and speed-quality trade-off.
Last verified: July 27, 2026. Product behavior was checked against the public desktop repository and linked model documentation.
Quick verdict
Use Large v3 Turbo as the multilingual starting point on Mac or Windows. Use Base (English) when download size and speed matter more than recognition quality, or Small (English) for a middle ground. On an Apple Silicon Mac, try Parakeet V3 when you want the fastest on-device option and your language is in its supported set. Choose Large v3 only when you can accept the largest download and slower processing.
These are starting points, not universal accuracy rankings. Microphone, accent, vocabulary, noise, hardware, and recording length all affect correction effort.
Shipped local models
| Model shown in Voicetypr | Download | Platforms | Language scope | Good starting role |
|---|---|---|---|---|
| Base (English) | ~148 MB | Mac + Windows | English only | Smallest download and fast English drafting |
| Small (English) | ~488 MB | Mac + Windows | English only | Middle ground for English |
| Large v3 Turbo | ~1.62 GB | Mac + Windows | Multilingual Whisper | Best default balance for multilingual use |
| Large v3 | ~3.10 GB | Mac + Windows | Multilingual Whisper | Maximum shipped Whisper quality when speed and storage are secondary |
| Parakeet V3 | ~500 MB | Apple Silicon Mac only | 25 documented languages | Fast multilingual on-device dictation on supported Macs |
| Parakeet V2 (English) | ~480 MB | Apple Silicon Mac only | English only | Fast English-only option on supported Macs |
The download sizes and availability above come from the current public Voicetypr model implementation. Voicetypr's in-app speed and accuracy badges are product guidance, not results from a published hardware-neutral benchmark.
Choose by platform
Windows 10 or 11
Choose a Whisper model. Parakeet is not shipped on Windows because Voicetypr's Parakeet path uses Apple's Neural Engine. Start with Large v3 Turbo if your machine has enough storage; step down to Small (English) or Base (English) when responsiveness matters more.
See the Windows installation guide and offline Windows buyer guide for setup and built-in alternatives.
Apple Silicon Mac
You can compare both Whisper and Parakeet. Start with Parakeet V3 for a fast supported-language path, or Large v3 Turbo when you need Whisper's broader multilingual coverage. Test the same short recording with both and compare the edits you actually make.
See the Mac installation guide before downloading a model for the first time.
Intel Mac
Choose a Whisper model. Parakeet requires Apple Silicon in the current implementation. Start small, then move up only if the correction improvement is worth the added download and processing time.
Choose by language
Base (English), Small (English), and Parakeet V2 are English-only. Large v3 and Large v3 Turbo use multilingual Whisper models. Parakeet V3 currently lists:
English, Spanish, French, German, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Swedish, Russian, and Ukrainian.
Language support does not prove equal quality across languages, accents, or specialist vocabulary. Voicetypr has not published a reproducible per-language benchmark, so test names, punctuation, numbers, jargon, and code-switching before committing to a model.
A five-minute model test
- Record the same 30–60 second passage for each candidate model.
- Include names, numbers, punctuation, and the specialist terms you use at work.
- Keep the microphone, room, and distance unchanged.
- Count meaningful corrections rather than judging how polished the first line looks.
- Choose the smallest or fastest model that keeps correction effort acceptable.
Do not compare different passages or silently clean one output first. This test helps you choose on your hardware; it is not a vendor-neutral benchmark.
Local, cloud, and remote are different sources
This guide covers local models stored on your machine. Current Voicetypr builds can also expose connected cloud transcription providers and another Voicetypr device on your local network. Selecting a cloud provider changes the data path because audio is sent to that provider for transcription.
Read Voicetypr privacy and data handling before enabling a cloud source. For a wider market comparison, see best offline dictation software.
When you are ready to test the model on your own hardware, download Voicetypr. The three-day trial does not require a card.
Sources and methodology
Voicetypr publishes this guide. Model names, sizes, platform filters, and language lists were checked against the public desktop source on July 27, 2026. No WER, latency, or RAM number is presented because this update did not run a controlled benchmark across supported machines.