Why does Whisper sometimes return words that are not even true words.

[RndyP]

I'm building my own "error dictionary", that converts single words returned by Whisper, that are not English. (I'm using ggml-base.en.bin) I get words that make no sense. This seems to occur most frequently when I am using Whisper as a command processor, where I speak a single word. I get these words in a "clean" environment, that is, there is no background noise. Here's an example:

This is my error dictionary entry for "compass"

	{{"compass"}, <---- the word i want
	The wrong words I may get:
            {"cropus","crappus","karpas","copper","coppish","copies","corpus","carpus","campus","cappas"
	,"carvus","coppers","calculus","carpers","crampus","carpets","carpos","cabas"}},

Some of these words I cannot even find in a Google search. Is this because the huge but weaky trained Hugging Face models contains these tokens?

[misutoneko]

AFAIK the original training material is kept under wraps, but it might be that the word "compass" just isn't sufficiently presented there.
And there's also the possibility of proper names...I mean any utterance can also be a place name, or some person's name.
That's something these models struggle with a lot sometimes (but can be fixed by prompting).

EDIT: Just ran some tests, "compass" was transcribed just fine when it was a part of a speech. And come to think of it, there's the word "compassion" which is more common and probably uses the same tokens. So the word's in the model alright, but isn't always detected when it's standing on its own (bigger models fare better in this regard, of course).

[RndyP]

I see the same thing. If I use "compass" in a sentence, it seems to work better. It understands context.
However, I find it hard to believe that "crappus" and "cropus" are more popular than "compass". I guess it is what it is.