Differences in large-v1, v2 v3 models?

[ItsNoted]

What are the main differences in large-v1, v2 and v3 models? They all seem to be nearly the same exact size so I am curious how I can see what the differences are?

[jhj0517]

Hi. Simply trained differently.

Details are here:

• Announcing the large-v2 model openai/whisper#661
• `large-v3` release openai/whisper#1762
• `turbo` model release openai/whisper#2363

They give you different results for the same input.

I haven't run any GOOD benchmarks in terms of WER (Word Error Rate, Lower is better.) between like large-v1 vs large-v2 etc, so here's my personal experience:

large-v1 is the first large model and in most cases it gives a worse result than large-v2.

A bit controversial is large-v2 vs large-v3.
In my personal experience, large-v3 often causes really bad hallucinations if the audio has a little bit of noise.
( See #152 (comment) and openai/whisper#2378 for more info )

But if the audio is really clean without much noise, like an ASR benchmark dataset, large-v3 will give you more accurate timestamps in my experience.

And if you don't care much about accurate result, you can consider using large-v3-turbo because it's lighter, faster, with really minor result quality downgrade than large-v3.

You can see how to use it in the Web UI at #309 (comment).

TL DR;
If your audio is clean (no noise), use large-v3.
If not, use large-v2.
If you're OK with a slight loss of quality compared to large-v3, use large-v3-turbo for faster transcription.