Earkeep
Languages and models
Earkeep transcribes locally with Whisper or Parakeet models, hardware-accelerated on Apple Silicon. Whisper covers 99+ languages. Parakeet-TDT v3 covers 25 European languages, including Czech, and is the faster of the two. Both run entirely on your device: no audio is ever sent anywhere to be transcribed. The full language list is below, marked with which of the two models covers each one, so you can check your own language before you buy.
Which model should I use
Mostly English meetings: use Parakeet-TDT v3. It tops the open ASR leaderboard on speed and beats Whisper large-v3 on English accuracy, so it's the faster, more accurate choice when English is your primary language.
Any other language, or a mix of languages: check the language list below. If your language is one of the 25 Parakeet covers, it's the faster option there too. If it isn't, or your meetings switch between languages Parakeet doesn't cover, use Whisper: it covers far more ground, at the cost of being slower for a given model size. Slower is worth putting in proportion, though: Whisper still transcribes as people speak, so lines land on the timeline while the meeting is still going. What you give up is headroom, not the live transcript.
Either way, you're never locked into one choice. The model is a setting you can change at any time in Earkeep's Settings → Transcription, and a separate transcription-language setting there controls which language Whisper decodes toward, or auto-detects if you leave it unset.
Model comparison
| Parakeet-TDT v3 | Whisper | |
|---|---|---|
| Languages | 25 European languages | 99+ languages |
| Speed | Tops the open ASR leaderboard on speed | Varies by size; smaller sizes are faster, larger sizes are slower |
| English accuracy | Beats Whisper large-v3 | Best accuracy at large-v3, the largest size |
| Sizes offered | One size (int8-quantized) | Six sizes, tiny through large-v3 |
Every model Earkeep offers, what each one is for, and what it costs to download once:
| Model | Size | What it's for | Approximate download |
|---|---|---|---|
| Whisper | tiny | The fastest and smallest, and the least accurate with it | 75 MB |
| Whisper | base | Still small; a step up in accuracy from tiny | 142 MB |
| Whisper | small | The point where accuracy starts being good enough to trust unread | 466 MB |
| Whisper | medium | More accuracy again, for a noticeably slower decode | 1.5 GB |
| Whisper | large-v3-turbo (default) | A distilled large-v3 that stays close to it on accuracy while decoding several times faster, which is what makes a near-top-tier model practical all day | 1.6 GB |
| Whisper | large-v3 | The accuracy ceiling of the family, in particular for speech that switches language mid-sentence, at the cost of size and speed | 3.1 GB |
| Parakeet-TDT v3 | One size (int8-quantized) | NVIDIA's model for 25 European languages: tops the open ASR leaderboard on speed and beats Whisper large-v3 on English accuracy, inside that language list | 670 MB |
On-device, offline, downloaded once
Whichever model you pick, the mechanics are the same:
- Inference runs on-device, hardware-accelerated on Apple Silicon. Audio never travels anywhere for the transcription step, and transcription keeps working with no network connection at all.
- Models download once and run offline after that. Earkeep fetches the model files you choose the first time, stores them on your disk, and doesn't need to re-download or phone home to keep transcribing.
- You can change models at any time in Settings → Transcription, and Earkeep downloads the new model the first time you switch to it.
- A transcription-language setting exists in-app, separate from the model choice: it tells Whisper which language to decode toward, or you can leave it on auto-detect.
See private by design for the full mechanism behind "audio never leaves your device," including why on-device inference is what makes that claim true for transcription specifically, not just for storage.
The complete language list
The 99 languages Whisper is trained to transcribe, available at every Whisper size Earkeep offers. The 25 set in bold are also covered by Parakeet-TDT v3, so for those you can run either model and Parakeet is the faster of the two. The rest are Whisper only.
- Afrikaans
- Albanian
- Amharic
- Arabic
- Armenian
- Assamese
- Azerbaijani
- Bashkir
- Basque
- Belarusian
- Bengali
- Bosnian
- Breton
- Bulgarian
- Burmese
- Catalan
- Chinese
- Croatian
- Czech
- Danish
- Dutch
- English
- Estonian
- Faroese
- Finnish
- French
- Galician
- Georgian
- German
- Greek
- Gujarati
- Haitian Creole
- Hausa
- Hawaiian
- Hebrew
- Hindi
- Hungarian
- Icelandic
- Indonesian
- Italian
- Japanese
- Javanese
- Kannada
- Kazakh
- Khmer
- Korean
- Lao
- Latin
- Latvian
- Lingala
- Lithuanian
- Luxembourgish
- Macedonian
- Malagasy
- Malay
- Malayalam
- Maltese
- Maori
- Marathi
- Mongolian
- Nepali
- Norwegian
- Nynorsk
- Occitan
- Pashto
- Persian
- Polish
- Portuguese
- Punjabi
- Romanian
- Russian
- Sanskrit
- Serbian
- Shona
- Sindhi
- Sinhala
- Slovak
- Slovenian
- Somali
- Spanish
- Sundanese
- Swahili
- Swedish
- Tagalog
- Tajik
- Tamil
- Tatar
- Telugu
- Thai
- Tibetan
- Turkish
- Turkmen
- Ukrainian
- Urdu
- Uzbek
- Vietnamese
- Welsh
- Yiddish
- Yoruba
Related
- Private by design: why audio never leaves your computer, and never touches disk, for either model.
- Always-on capture: why a fast model matters when transcription runs all day.
Frequently asked questions
Does Earkeep support my language?
Almost certainly, through Whisper, which covers 99+ languages (full list above). If your language is also one of Parakeet-TDT v3's 25 European languages, you can choose Parakeet there for extra speed and, for English, better accuracy than Whisper's own top model.
Can I switch models after installing?
Yes. The model is a setting, not a one-time install choice. Change it any time in Settings → Transcription; Earkeep downloads the new model the first time you switch to it and keeps using the previous one until then.
How big are the model downloads, and do I need to keep re-downloading them?
Whisper ranges from 75 MB (tiny) up to 3.1 GB (large-v3); Parakeet-TDT v3 is about 670 MB. Each model downloads once and then runs entirely offline: Earkeep doesn't re-fetch it or phone home to keep transcribing with a model you already have.
What accuracy should I expect?
It depends on the model and size you pick. Whisper's large-v3-turbo (the default) stays close to full large-v3 accuracy while decoding several times faster; large-v3 itself is the accuracy ceiling of the Whisper family, especially for mixed-language speech. Parakeet-TDT v3 beats Whisper large-v3 specifically on English accuracy, while topping the open ASR leaderboard on speed, but only covers its 25 languages.
Does my audio go anywhere during transcription?
No. Transcription runs on-device, hardware-accelerated on Apple Silicon, whichever model you choose. Audio is processed only in memory and is never uploaded, never sent to an API, and never written to disk. See private by design for the full mechanism.
Find your language, then try it
14-day full trial, no account required.