Whisper generates incorrect subtitles

[carlos-truong]

Hi,
I'm trying to transcribe this video:
'Hide N' Seek': A Social Deduction Game Using Controllable Language Models

Here's the download link.

But the subtitle it generates is incorrect. I’m currently using:
Model: small.en
Version: v20250625

I can still transcribe other videos using Whisper. So I'm not sure what's wrong with this video? Is there any way to fix it?

[carlos-truong]

Can anyone help?
Here’s the transcribed text, but it's wrong.

form after the
I appreciate the opportunity to ask questions
about program requirements here to the extent
that our gains apply to this particular topic.
So if you're in Africa, like our gently sold in
correct or inflection correct?
So a lot of partners out there
are sitting around just the Atlantic cost
and they really sit down.
So they just start out reporting
as well as some of their children
will be VWERS in Rather Who will ."
about as a worked giant.
He gives their calls weke energizing
at the faculty who see it
as more and more aggressive.
And we're going to go all the way
towardshouling the alternative
to how they're actually saying it's
more and more effective.
That's an idea that was first
Thank you,
which means that when you're about to do the last number of times,
the rest of the time is going to be the same.

Hide N Seek - A Social Deduction Game Using Controllable Language.txt

[misutoneko]

There seems to be a lot of background noise which Whisper likes to interpret as speech, so maybe filtering those out would help some.
That said, I've noticed that some samples simply require a large model, the smaller ones won't do.

EDIT: All right, I see it now (or rather, hear).
The artifacts are caused in the conversion from stereo to mono.
I guess you could process the channels separately (but maybe there's a better way?)

[carlos-truong]

The artifacts are caused in the conversion from stereo to mono.
I guess you could process the channels separately (but maybe there's a better way?)

Thank you for the reply!
I'm new to this. How can I process the channels separately?

[misutoneko]

Hi, I can't remember the exact commands that I used to extract the channels, but probably ffmpeg and sox were involved.
So maybe something like:

ffmpeg -hide_banner -y -i Hide_n_Seek.mp4 -ar 16000 -c:a pcm_s16le stereo.wav
sox stereo.wav channel_1.wav remix 1

That'll only extract one of the channels and you'll get a mono wav file as output, but I'm not sure about the numbering scheme here. Some sources say the numbering starts with 1 but sox seems to accept 0 as well...
There are many other ways to do this, actually I think ffmpeg can do it all by itself too (without sox) but the command line gets a bit complicated if done that way.

The bigger question of course, is if this kind of thing can be detected somehow?
Normally it shouldn't be necessary to extract the channels in this way.

[carlos-truong]

Thank you!
I’ll try it. I think it would be better if Whisper could detect and process all channels at the same time so we wouldn’t have to separate and process them individually like that.