of course translating to english works better

[00:00.000 --> 00:02.740]  Natural Examination from the English language
[00:02.740 --> 00:04.720]  The basic level
[00:04.720 --> 00:08.640]  You will hear two times the texts from the first to the third
[00:08.640 --> 00:11.720]  Before listening to each text, you will hear the sound
[00:11.720 --> 00:13.320]  The sound
[00:13.320 --> 00:16.800]  In the recording, there are interruptions to get acquainted with the commands
[00:16.800 --> 00:19.600]  and the content of the task is signalized by the sound
[00:19.600 --> 00:21.040]  The sound
[00:21.040 --> 00:24.160]  Unlock the individual tasks during listening to the recording
[00:24.160 --> 00:28.160]  and during the break after listening to them
[00:28.160 --> 00:30.600]  The first task
[00:30.600 --> 01:00.560]  Read the command and get acquainted with the content of the task

this also illustrates another quirk, that being how sometimes whisper just decides to not do punctuation and sticks with it

i think that could be helped with the

prompt

option

give it a sample of what we want the style to be, e.g.

Doctor number five is gonna like, fix the, the thing.

prompt engineering in speech recognition what has the world come to

ok so

prompt

doesn't do anything

prefix

though makes it go from

15. Burger King Foot Lettuce The last thing you'd want in your Burger King burger is someone's foot fungus, but as it turns out, that might be what you get.

to

Number 15, Burger King foot lettuce. The last thing you'd want in your Burger King burger is someone's foot fungus, but as it turns out, that might be what you get.

which i think is a decent improvement

@mega b prompt engineering real

i will note that

prefix

seems to be for when the text is in fact represented in the audio, since my more targeted prompt

Number twelve. Number thirteen. Number fourteen.

made it skip over number fifteen entirely

update: can be transcribed with the stutters by adding

uh, um, like, well,

but it adds a ghost

this is,

at the start

never thought i'd enjoy an autoregressive model

Cools, but what sucks about prompting is that it costs some tokens to be used. It's obviously still useful tho, so it's probably worth endeavoring to conclude the most efficient prompts.

there's a token cost?

or do you mean because of attention

or

bhh words

--- i wonder if we could do a softprompt

barring that here's what a human-engineered prompt should include: - an unabbreviated title, e.g.

Doctor

,

Miss

etc - a number word - punctuation (maybe even ungrammatical-but-matching-intonation) - filler words - a swear word (there's at least one reported instance of it doing censorship but it seems to be rare) - be unlikely to be the start of a sentence because attention could mess things up

I could be wrong but that might be why they cut to 30 seconds audio automatically

I haven't seen censoring, I've seen the f word in rap God and "motherfucker" a ton

I'm sure YouTube censors tho

yeah i'm just going off of one reddit comment

Maybe they scraped a tiny bit of yt

to which the immediate reply was "haha closedai puritans"

😔

pff i accidentally put an instrumental through it and the hallucination is real

it should've outputted musical notes

if only they trained it on the subtitles for

tom scott plus

actually hm a softprompt would be better anyway because it could affect things like how often it splits the audio (for link2model 2)