Best for most people
NVIDIA Parakeet V3
Fast, highly capable, and multilingual. The recommended default for 25 European languages.
- Languages
- 25 European languages
- Approx. size
- ~483 MB
Local models
Choose the right local speech-to-text model
VoicePipe offers four speech-to-text models that run locally on your Mac. Choose by language, speed, and storage — no model expertise required.
Parakeet, further trained
Parakeet Ultra
Moondream’s further-trained version of Parakeet V3. Try it for the same 25 languages with a larger download.
- Languages
- 25 European languages
- Approx. size
- ~632 MB
Lightweight English
NVIDIA Parakeet 110M
The smallest model in VoicePipe, built for fast, lightweight English transcription.
- Languages
- English only
- Approx. size
- ~227 MB
Widest language coverage
Whisper Large V3 Turbo
Choose Whisper when you need one of the many languages not supported by Parakeet V3.
- Languages
- 100 languages
- Approx. size
- ~648 MB
Language Finder
Find the right model for your language
Search by language to see which VoicePipe model is the best fit. Multilingual models detect the spoken language automatically.
A language may be completed inline. Press Right Arrow at the end of the text or Enter to accept the suggestion.
Start typing a language to get a recommendation.
Browse all 100 supported languages
Parakeet V3 languages
25- Bulgarian
- Croatian
- Czech
- Danish
- Dutch
- English
- Estonian
- Finnish
- French
- German
- Greek
- Hungarian
- Italian
- Latvian
- Lithuanian
- Maltese
- Polish
- Portuguese
- Romanian
- Russian
- Slovak
- Slovenian
- Spanish
- Swedish
- Ukrainian
Additional Whisper languages
75- Afrikaans
- Albanian
- Amharic
- Arabic
- Armenian
- Assamese
- Azerbaijani
- Bangla
- Bashkir
- Basque
- Belarusian
- Bosnian
- Breton
- Burmese
- Cantonese
- Catalan
- Chinese
- Faroese
- Filipino
- Galician
- Georgian
- Gujarati
- Haitian Creole
- Hausa
- Hawaiian
- Hebrew
- Hindi
- Icelandic
- Indonesian
- Japanese
- Javanese
- Kannada
- Kazakh
- Khmer
- Korean
- Lao
- Latin
- Lingala
- Luxembourgish
- Macedonian
- Malagasy
- Malay
- Malayalam
- Māori
- Marathi
- Mongolian
- Nepali
- Norwegian
- Norwegian Nynorsk
- Occitan
- Pashto
- Persian
- Punjabi
- Sanskrit
- Serbian
- Shona
- Sindhi
- Sinhala
- Somali
- Sundanese
- Swahili
- Tajik
- Tamil
- Tatar
- Telugu
- Thai
- Tibetan
- Turkish
- Turkmen
- Urdu
- Uzbek
- Vietnamese
- Welsh
- Yiddish
- Yoruba
At a glance
Four models, compared.
Start with Parakeet V3, try Ultra as an alternative for the same languages, or choose by download size and language coverage.
| Model | Best for | Languages | Relative speed | Transcription quality | Approx. size |
|---|---|---|---|---|---|
| NVIDIA Parakeet V3 | Most people | 25 European languages | Fast | Strong results in the July 2026 benchmarks | ~483 MB |
| Parakeet Ultra | An alternative to V3 | 25 European languages | Fast | Lower error rates than V3 in Moondream’s benchmarks | ~632 MB |
| NVIDIA Parakeet 110M | Lightweight English | English only | Fast | Strong English benchmark performance | ~227 MB |
| Whisper Large V3 Turbo | Languages outside Parakeet | 100 languages | Balanced | Capable multilingual transcription | ~648 MB |
NVIDIA Parakeet V3
- Best for
- Most people
- Languages
- 25 European languages
- Relative speed
- Fast
- Transcription quality
- Strong results in the July 2026 benchmarks
- Approx. size
- ~483 MB
Parakeet Ultra
- Best for
- An alternative to V3
- Languages
- 25 European languages
- Relative speed
- Fast
- Transcription quality
- Lower error rates than V3 in Moondream’s benchmarks
- Approx. size
- ~632 MB
NVIDIA Parakeet 110M
- Best for
- Lightweight English
- Languages
- English only
- Relative speed
- Fast
- Transcription quality
- Strong English benchmark performance
- Approx. size
- ~227 MB
Whisper Large V3 Turbo
- Best for
- Languages outside Parakeet
- Languages
- 100 languages
- Relative speed
- Balanced
- Transcription quality
- Capable multilingual transcription
- Approx. size
- ~648 MB
Included with VoicePipe
Download only what you need
All four transcription models are available without a license for free file transcription. Download and switch between them inside the app. Recordings and transcripts stay on your Mac during transcription.
Methodology and sources
Our quality guidance uses the Hugging Face Open ASR Leaderboard snapshot from 10 July 2026, covering its English benchmark and five tested European languages. The linked leaderboard may reflect newer results and methods. The benchmark compares upstream model implementations; VoicePipe runs converted Core ML artifacts, so on-device results can differ by language, accent, microphone, recording conditions, and runtime. Ultra’s quality guidance comes from Moondream’s separate September 2026 comparison with V3, not the July leaderboard. It does not establish a ranking across all four models. Relative speed labels are product guidance; the existing VoicePipe-on-Mac benchmark does not include Ultra.
All local models included
Your voice. Your language. Your Mac.
Choose a model in the app and turn speech into text without sending recordings to a cloud transcription service.
Buy VoicePipe