Whisper supported languages: the full list

Whisper large-v3 and its Turbo version recognise 100 languages. Here is the full list, with each language's own name and the code you pass to Whisper, and a mark on the 25 that NVIDIA's Parakeet engine also covers.

The list is not copied from a blog post. It is generated from the language table inside whisper.cpp — the same table Vocula's language picker reads — so it matches what the model actually accepts. That is one more than the 99 you will see elsewhere: large-v3 gave Cantonese (yue) its own code, and lists built from the earlier models do not have it.

“Supported” is not the same as “good”

Whisper will attempt every language below, but accuracy varies widely. OpenAI’s own paper ties each language’s error rate to how much training audio it had, and the languages with little data score far worse than the widely spoken ones (Radford et al., 2022). If your language is near the bottom of the training data, test it on your own voice before relying on it.

The other failure is detection. Left on automatic, Whisper chooses the language from the first seconds of speech, out of all 100, and can pick one you never speak. Vocula asks which languages are yours and only lets it choose among those — how that works.

All 100 languages

Whisper or Parakeet

Parakeet covers 25 European languages, downloads as 468 MB instead of 1.62 GB and returns text faster. Whisper covers everything above and handles several languages inside one sentence better. The two engines compared.