What's the difference? (language="en" vs "medium.en" model)

[SimpleVictor]

Lets assume I only care about english. Can someone help me understand the difference between the 3 approaches below?

1. Using medium model with language="en" attached

whisper english-only.wav --model medium --language english

2. Using medium.en model with NO language="en" attached

whisper english-only.wav --model medium.en

3. Using medium.en model with language="en" attached

whisper english-only.wav --model medium.en --language english

[glangford]

Without language= whisper will perform language detection using the first 30s of audio. If you only care about English, then specify language=en to skip language detection.

medium is a multilingual model, whereas medium.en is an English only model.

From the README at https://github.com/openai/whisper#available-models-and-languages

The .en models for English-only applications tend to perform better, especially for the tiny.en and base.en models. We observed that the difference becomes less significant for the small.en and medium.en models.

See also the discussion here:

• Most accurate model for English? #1540