Saypane
English

Which speech model for which language, and how fast

Saypane runs open speech recognition models on your PC. Which one it uses depends on the language you speak and on how much processing power your PC has. Here is how it decides, with the speeds we measured.

Which model Saypane uses

LanguageDictationFile transcription
The 25 Parakeet languages (English, German, French, Spanish, Portuguese, Italian and 19 more)ParakeetParakeet
Other languages, freeWhisper smallWhisper small
Other languages, Pro, with a Vulkan GPU or 8+ CPU threadsWhisper large v3 turboWhisper large v3 turbo
Other languages, Pro, on a slower PCWhisper smallWhisper large v3 turbo (slower)

Parakeet TDT 0.6B v3 covers Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish and Ukrainian. It does not cover Japanese, Chinese or Korean; Whisper does.

How fast

Measured on our Windows test PC:

  • Parakeet: about 13 times faster than real time on 4 CPU threads. A 10-second dictation is ready in under a second; an hour of audio takes under 5 minutes. No GPU needed.
  • Whisper large v3 turbo: 0.6 to 1.0 times real time on the CPU alone, and about 1.9 times with Vulkan on integrated graphics. Fine for files if you can wait, too slow for comfortable dictation without a GPU or a CPU with 8 or more threads.
  • Whisper small: much faster than turbo, but noticeably less accurate. For Japanese in our test, it made about three times as many errors as turbo.

Your PC will differ. A recent desktop processor or a dedicated graphics card is faster; an older laptop on battery is slower.

Minimum hardware

  • Windows 10 or 11, 64-bit.
  • A processor with 4 cores and AVX2, for example an Intel Core of the 4th generation or newer, or any AMD Ryzen. Some low-end Pentium and Celeron chips lack AVX2.
  • 8 GB of memory. The models stay loaded while Saypane runs, so the first word is never lost to a loading delay.
  • Disk space for the models you use: 669 MB for Parakeet, 574 MB for Whisper turbo, 190 MB for Whisper small (included).

GPU acceleration

Saypane uses Vulkan, which works with NVIDIA, AMD and Intel graphics, including the integrated graphics in many laptops. You do not need an NVIDIA card or CUDA. If no suitable GPU is found, Saypane runs on the CPU.

Japanese, Chinese and Korean dictation

These languages use Whisper. Without a GPU, Saypane falls back to Whisper small for dictation, so you get text quickly but with more mistakes. For accurate Japanese, Chinese or Korean dictation, use Saypane Pro on a PC with a Vulkan GPU or a CPU with 8 or more threads. File transcription can always use large v3 turbo in Pro; it just takes longer on a slow PC.

Downloads

Whisper small and the Silero voice activity detector are included in the app, so Saypane works offline from the first start. The other models download when you first need them, from models.saypane.com or, as a fallback, from Hugging Face. Downloads resume after an interruption and are verified by checksum. Full model list.

Questions

Do I need an NVIDIA graphics card?

No. Saypane uses Vulkan, which also runs on AMD and Intel graphics, and it works without any GPU. For the 25 Parakeet languages, a 4-core CPU is enough.

Can I use my own Whisper model files?

Not in this version. Saypane uses the models listed above, which it has tested.

Last updated October 27, 2026. Still stuck? Contact support.

Other guides