Skip to content

Local speech recognition: Whisper and Parakeet measured

Converting speech to text can be done entirely on your own hardware, with open models such as Whisper and Parakeet. We measured how well and how fast that works.

In short

Calls, dictation, meetings: converting speech to text is one of the most useful AI applications. It does not have to go through the cloud. Open models run on your own computer, so recordings never leave the company.

The models

  • Whisper (large-v3) is an open model that understands many languages, including Dutch, English and German. Very accurate, somewhat slower.
  • Parakeet is a smaller, very fast model, strongest in English.

What we measured

We tested speech recognition on our own test recordings, on a Mac with Apple Silicon:

Model Word errors Speed
Whisper large-v3 0% in Dutch, English and German 7 to 9 times faster than real time
Parakeet 3.3% in Dutch, 0% in English and German 35 to 42 times faster than real time

So an hour of recording is transcribed in under ten minutes with Whisper, and in a few minutes with Parakeet. One caveat: this is a small test set; with a lot of background noise or dialect, the error rates are higher.

One surprise: first "cleaning up" a recording by separating music and voice made the recognition worse. The error rate rose from about 6% to 15 to 21%.

What you can use it for

  • Meeting minutes and call notes, then summarised by a language model.
  • Dictating instead of typing.
  • A voice assistant that first checks whether it has understood you correctly.

What to watch out for

  • Choose per language. In our test, Whisper made fewer word errors in Dutch. For English, both transcripts were error-free and Parakeet was faster.
  • Have important texts proofread, especially names and numbers.
  • Ask for permission before you record conversations.

Would you like advice on local speech recognition? Request a free advice conversation.

Updated on

AI

Personal data in the AI cloud: who is watching?

A conversation with an AI service in the cloud is not private by default. A case in Canada shows what a provider can see and share. What does that mean under the GDPR, in the United Kingdom and in Switzerland?