Transcribing Videos Using a Mysterious New Technology
HTML-код
- Опубликовано: 8 сен 2024
- Just like IT guys everywhere, hand-typed subtitles aren't appreciated until they go wrong, and then they're hated more than were there no subtitles at all. This is one job I don't mind AI taking sooner rather than later.
Here's the really cool one online that's great and stuff: huggingface.co...
The one on my PC made 3-4 mistakes with this video, but it's still really good.
People missing the point that the AI-generated subtitles used in this video represent unedited AI-generated subtitles. They're a remarkable improvement over using the original script or RUclips's own subtitles. Mistakes with these are so few and far between I might even be able to correct these in review in real-time.
If nothing else then maybe it will give you extra appreciation for actual human-edited subtitles
it's just ironic how the only error present in this video changes the entire meaning of the sentence, rather than being a little typo! a little bit more luck and we wouldn't have noticed a thing, which brings me to my next point: when you do something well, no one notices, but when you screw something up, it's painfully obvious. i know it's discouraging, but please know that the people who do use your subtitles often (like me) appreciate them very much, and we do notice that little extra effort that most others don't make, it's just not something we think about pointing out! so please keep going! it's a dull, thankless task, but you have no idea how grateful some of us are for it
The error is so easy to hear so I wouldn't even bother spending time fixing it for the 3 people who watch videos with subtitles turned on, it seems good enough.
@@mushroomcrepes4780 (please don't forget about deaf people who can't hear the mistake. subtitles are for them, too!)
It's not a huge improvement over RUclips's captions lol, there are more mistakes and worse ones. But with punctuation etc so easier to read, and certainly a better version to edit from.
The reason I oppose auto generated subtitles is because it's just fucking lazy, and for people who genuinely need them for accessibility, having the incorrect terms makes a video worse at best, and worst case scenario, makes the video entirely unwatchable. The vast overwhelming majority of people who use auto generated subtitles DO NOT EVER look for any errors and just slap them on the video as is. So you might be right that this is a good tool, but it makes subtitles worse for the people who actually need them because the vast overwhelming majority of people who use it won't ever fix its errors.
...but can AI pass the "hell of a lot" test?
This is exactly what I was going to ask.
I bet it can't get the usermenchbark right too
The truest of questions
What does that mean? I'm curious
@@kortyzol420 Phillip has a vid where he tries to say "hell of a lot" and fails multiple times, it's quite funny and he ended up referencing it from time to time when he says hell of a lot (not many times)
2:30 Written as "free. It's not really good" instead of freakishly good
The youtube auto-generated subs transcribed that correctly, interestingly enough.
The irony writes itself. It's kinda funny in a way. lol
lmaoo that's incredible
And the irony of RUclips's built-in captions getting that line right.
@@noth1ngnss921A broken clocks right atleast twice a day
I can't believe AI generated Philip was being forced by real Philip to hand type all of the transcriptions for his videos
it's interesting that sometimes the small transcribing errors might completely change the message, like in 2:28
Thankfully these errors are so few in between that sentences around error usually give enough context that you can deduce what went wrong. Confused for a little while but you should be able to get it right. And if creator reads over it at least once its easy to correct since its so glaringly obvious error.
It's to be expected with AI. The AI is trained on data of speech that makes sense, so it'll make an effort to generate text that makes sense. So it'll generate text that makes sense even if it makes a mistake. This could be seen as a downside, as it makes mistakes more difficult to spot.
there's definitely some errors in the subtitles but its not catastrophic
I know the automated subtitles are generally pretty good, but there's one thing that automation can't quite beat out hand-made subtitles on, and that's pacing. Actual hand-made subtitles tend to have proper pacing in them, usually starting a new subtitle when a new sentence is spoken, having breaks where it feels natural, and not revealing information until it's necessary, such as a surprise that comes later in the video. Plus, if you have multiple people, you can say who's speaking at what point with handmade subs, can't do that with automated ones yet. It also screwed up at 2:31, transcribing "freakishly good" as "free. It's not really good".
He already wasn't doing that by hand, as he pasted the entire transcript into RUclips and let it set the timings. But if you have the time then manually edited ones can be even better yes
yeah, like how a lot of subtitles will end a line with a dash if a voice gets cut off or is about to reveal a large piece of information, so that you can't read ahead and spoil yourself before the video reveals it.
say for instance in a crime drama where a character is about to reveal who the killer is, the handwritten subtitles might read:
"and this is why I believe the killer is-"
[dramatic pause]
"Jeff the Mailman!"
whereas automatic ones might read
"and this is why I believe the killer is Jeff the Mailman!"
which completely ruins the pacing of the scene for people using subtitles
@@speedstyle.Only time I've ever done manual timings was with the Odgrub videos and whenever a song plays and the old subtitle would remain on-screen for an irritating amount of time. And yes, it takes a very, very long time, especially with RUclips's fiddly subtitle timings editor.
@@kliksphilip I use SubtitleEdit, which is open source and has a Whisper plugin built in. It doesn't have auto-timing like YT so if you're starting from the script it's less useful, but Whisper itself can return timestamps as srt. It also has a side-by-side mode which is great for translation but that's not really your workflow
@@kliksphilipI tried Aegisub and Kainote for frame by frame timing editing if you're looking for alternatives.
RUclips's built-in editor is so lackluster it's no wonder they said nobody made community subtitles when killing off the feature.
congratulations philip you've turned your comment section into a : 2:30
Just for the record i think I was first
I work on a cloud transcription service, I always watch your AI videos, but this was quite surreal to hear you talk about something I do every day.
I think the auto-generated ones from RUclips are still fairly good, especially in comparison to just like 1-2 years ago. It still makes mistakes and it's also based on what the video is about but I've also seen it transcribe "slang-words" that aren't really in any dictionary or similar.
true. i think the main reason people dunk on them so much is because for whatever reason they still haven’t done anything when it comes to formatting and punctuation
As a big, long time fan of you, I watch practically every video you upload shortly after their release. Thing is, this isn't always at a time when I can listen to it, so I turn on subtitles. I very rarely notice any large issues, and the fact you go through all the work to do it means a lot because it allows me to enjoy your content when and where I can. Thank you Philip!
My brother in christ, Whisper large v3 is offline too. the online version is basically just a preview.
V2 seems to work better than V3 a lot of the times though. (at least when transcribing songs it hallucinates a lot less and doesn't add stuff that weren't there as frequently) Also it's possible he is using like a distilled version of Whisper which is gonna perform worse probably than the full Large model.
The feedback you get on subtitles are like that one quote from Futurama "When you do things right people won't be sure you've done anything at all."
I've just been so glad you have subtitles. Accessibility aside, it also means I'm able to watch the video in a noisy environment without turning the volume up high and disturbing people around me (headphones would fix that, yeah, but I don't always have them on me)
This all reminds me of that Tom Scott clip where he says "(mocking) 'Oh look at me! I bought a Lamborghini!' Buy some damn subtitles!"
"transcribed feature online is free. It's not really good," 2:29💀
I wonder if that was on purpose.
>makes video promoting "super good" AI subtitle feature
>AI subtitle feature fucks up specifically at the part where he is praising it
And as always, manual subtitles win again.
@@potatosordfighter666Did you guys even watch the video? He was talking about another AI website which is the one he didn't use in this video because he preferred to use one that he could run locally, which is worse.
I really appreciate it when you work on subtitles, it is something thats very easy to take for granted,
but any time I see a video that uses man made subtitles it makes my life a lot easier.
I would definitively take my time to go through the scripts with the videos after generating them,
just to correct what the AI misheard, timing and structuring.
At the same time, the lack of subtitles on your videos spesifically is less of an issue for me because you're well spoken
and articulate your words clearly
Great example of AI simplifying menial labour! You are doing a fantastic thing by making your content more accessible. I am glad you found a way to achieve that more efficiently
Even though I'm not commenting about them, I always use subtitles when available and appreciate you as a youtuber for having subtitles on nearly every video. It's like someone else said here already, survivorship bias. You don't hear from people who have nothing to complain about subtitles, because well, they don't have a reason to complain.
I’ve always appreciated your subtitles and had massive respect for the amount of effort you put into them. Hopefully AI will save you from this dull dusk, although right now it’s not perfect, it’s damn near close.
As someone who has done subtitle translation for free and has typed out transcripts of video files of events etc. (from 9 hour long videos, sometimes) I totally understand the pain. The tool you shared is really cool by the way. The next time someone throws an hour long video at me to transcribe, I'll probably run it through this first :)
I really appriciate subtitles, as they can just be quite helpful in many scenarios. I actually use them for language learning with French - youtube's transcriber kinda sucks ass, especially with the "word per word" timing which is horrible to read, and especially since French has so many homonyms. Anyways, I applaud you for doing them manually in the past and now implementing this into your workflow, it's a great resource for many small groups :)
I feel like this can be affected by survivorship bias, since people who would normally be okay with your subtitles would have no reason to specifically point out that they're good, meanwhile when they're gone then they have a reason to comment about it being missing or inaccurate. It self-selects such that you'll pretty much always see negativity, so don't take it too hard on yourself.
I myself know this, since I'm hard-of-hearing, and using your subtitles along with many other people helped quite a bit. I even noticed right away when you had stopped doing captions for your videos, and you replied to one of my comments, explaining how youtube was slowly breaking support for this feature and combined with the negativity, you decided to stop doing it.
I'm not really as afraid of AI as other people, though i'm still cautious about it. I will say though that if you can go through with this AI feature, then i'm all for it. I've watched first-hand the developments of automated transcriptions, from back in the day when youtube subtitles would be hilariously innacurate, to nowadays with applications like Zoom where it's good enough that it'll even transcribe conversations happening in the background as i speak with a surprising level of precision.
Youtubing is a thankless job, so I'm thankful that you not only keep up such a nice standard of quality and dare-i-say professionalism that is rare nowadays, what with youtube continually pushing shorts and such. Kliksphillip is a nice channel that shows me the inner machinations of what being a youtuber is like, and the trials and tribulations that come with it, so as a closing remark, i'll say: Thank you Phillip, you're a pretty cool youtuber.
If you meant that the HuggingFace web UI was better than the local one, you can easily "duplicate the space" to not rely on *that* space, also skipping the queues. You will pay for GPU time though. Also, it should be possible to run that same model+config on your local machine. HF stuff is OSS
I use faster-whisper, which is about 3x faster than OpenAI's library with identical text output; and Colab notebooks to run online rather than huggingface/replicate/etc. There's also Nvidia Parakeet which is supposed to be better, but I prefer the speed
Yeah I've been pleasently impressed with Preimere Pro's AI subtitles. It's very effective and I usually just have to fix up a few words here and there. Adobe's AI stuff in their programs has been really neat. Photoshop's generative fill is extremely impressive. It's especially useful for me when I need to expand an image's aspect ratio to 16:9. Super cool
As someone who also types out their subtitles, I feel your pain. I was thinking about trying something similar but reading along with it afterwards to make sure its correct. Something I really hate is seeing AI subtitles baked into youtube shorts. I understand taking a shortcut on rarely used feature for a 10 minute video, but nothing says content farm like an AI subtitle error in a 10 second video.
Davincis auto subtitles is also pretty good. Its also useful for if you want to bake it into the video cause you can stylize it if you need to, and if you wanna add animations to them theres a free addon thatll convert them from subtitles to layers
Just make sure to at least check them.
Linus Tech Tips had some horrible errors in almost every video subtitle forever. Not even stuff like slang, but GPU will turn into CPU will turn into See Pee You.
I watch videos with subtitles a lot because my hearing isn't great and this kind of stuff is worse than no subtitles at all.
Cool feature is that Microsoft word has an auto transcribe feature where you can upload an audio file and it splits the text by speaker.
I've only used it once but it worked well enough to transcribe a audio-recorded interview.
I love the idea of using AI to transcribe subtitles. A couple of times, when I've wanted to make subtitles and was looking for good software that I could use for that purpose, I didn't find much. Most of it was old, and not really all that good. I also don't want to use youtube's own tool, because, just like you, I'm the type of person that wants to have everything happen locally, on my own computer. Also, it really is time consuming. However, as impressive as AI is, it still makes some mistakes (as I've written below), and would probably need to be trained some more before it's perfectly accurate. I also wanted to thank you for even bothering to make subtitles. It's really nice for if I for some reason can't listen to the audio (such as when I'm in a public space and forgot to bring my headphones) or for when I zone out for a second and miss what you've said, I can just read the subtitles.
Mistakes I've noticed:
1:30 "so see what you think of" is missing
Others have already pointed out 2:29
2:48, "I can (tell you) now"
watching the whisper and youtube subs side by side showed that at least for this video the youtube subs were still more accurate...
As far as I know whisper v3 large is the same core model online as well as when running it locally. It's all a question of how the surrounding software prepares the audio and with which parameters and chunk lengths and section overlaps and all that its being fed to the network. There are lots of options to configure as well. For example beam search gives significantly better results but it takes a bit longer. So just looking for other implementations that also use whisper v3 large might yield you results much closer to the ones you get from the online service.
Very interesting! Nice to see someone caring about subtitles and finding a way to make it fit for you :) thanks Philip!
Also nice to see a creator looking at somewhat niche pieces of feedback, not the "I like this" or "I don't like this" but thinking "what about the subtitles"
Good enough to follow the video along, definitely better than RUclipss own caption generator. But there are some easy to spot errors, including one that changes the meaning of the sentence: freakishly good got trandscribed as "free".
Your videos seem rather light-hearted and always somewhat personal, like you're just talking to your viewers. So any errors in subtitles I can easily gloss over (I hadn't expected you were manually transcribing your videos, too). I think what really brings me out are video essay-type videos, where some of the vocabulary is really deliberate and not common speech, and when their, now I know, *automated* method of transcribing makes an error, it's a lot more of a blemish to me--especially because the videos themselves are actually really high quality with really good production value.
An example is kaptainkristian's Futurama video. A fantastic video in its own right, but in the first 90 seconds, it mistakes "and it's rare" -> "and swear", "meme potential" -> "mean potential", possessive "its" -> contraction "it's", "just a" ->"just the", and "Simpsons clone" -> "Simpsons club". It really brings me out! But since your videos are a lot more personal and seem way more free-form, I think 95% correct subtitles are simply just a nice thing to have.
_“I'm surprised that RUclips's own transcribing is still is bad as it is”_ comparing the Whisper and RUclips subs on this video, YT's were arguably better. It made some mistakes (smarmy/SMY, timed/Tim, lopping/loing, transcribe/trans) but they were closer to what you said. Whisper contains a small GPT model and is prone to hallucinate entirely if it misses something (“for every 1 minute”/“for a minute”, smarmy/SMAMI, “so see what you think of the”/“the”, “I'm in”/“I've been”, “freakishly”/“free. It's not really”). It does do punctuation and better capitalization but that's almost about subtitles vs captions than speech recognition quality
I never write a comment about it and don't know about others, but I often use subtitles. Seeing them well made makes a video feel significantly higher quality, and that has usually been something iv noticed with your videos.
Hopefully voice to text in the near future will feel like a solved problem. Especially for phone keyboards!
Nothing beats the flair subtitles with proper formatting (think of changing size/color/place akin to NicoNicoDouga's comment-on-video), proper pacing and adding some extras into subtitles.
I’ve been using Whisper Ai for automated lipsyncing in some TikTok’s (which I broke the blender file of). The blender addon I use which does the lipsyncing does a fairly good job with only sound, but it can do an almost perfect job with a script and subtitle. Genuinely, since I have ADHD sitting down and writing what I’ve already spoken kills me.
The AI I got to use it tho handle memory pretty well, it doesn’t do the audio in one go but splits it into smaller chunks your PC can handle. Does mean it can be incredibly slow but you can get some great results using the larger models available on it.
With this technology, especially the fastest GPU accelerated versions of whisper, you can transcribe an entire channel and then linguistically analyse them.
I think you should stick to the online AI transcriber, for now. It was noticably better. Smarmy was SMAMI (the fuck even is that?) and there were a number of other glaring issues that would confuse the hellalalot out of someone without audio to help them connect the dots (Like, deaf people, or people watching this on a bus with no headphones. Niche watchcases, yes, but they do exist, and especially the disabled have always BEEN the MAIN use-case for subtitles). From what parts of the online transcription there were shown, it had basically none of these issues.
I understand the desire to keep every part of your workflow offline, but the tech doesn't seem to be there yet for one reason or another.
This might not be applicable to your scenario, but this is what I used to do when I was uploading vids and making a conscious effort to make good subtitles:
- upload only audio of my narration to RUclips
- wait for it to automatically generate subtitles along with timestamps
- download the .srt file of the automatically generated subtitles and import them to Vegas (think I used the Vegasaur plugin for this)
- go through the subtitles as the video is playing and change anything that RUclips subtitled incorrectly
As technology improves, hopefully these issues will subside. Being early as I am, I'm suprised by the amount of negative comments about this.
This is a situation in which a technology does ~80% of what a human can do in a small fraction of the time, thus making it worth it to the creator. However, the audience doesn't see the potential of such time efficient practices, only the deficit in quality.
I figure most people don't have the time to consider nuance, so stuff is either 'good' or 'bad' and AI is immediately considered one of those things without weighing up the pros and cons etc.
I remember back when RUclips suddendly declared they would be removing community-generated transcriptions from videos, apropos of nothing other than perhaps believing contributors had very little to gain since it's the RUclipsr in question who's benefitting from them without even getting involved in the process, and first thought I had was precisely "how come Nvidia has already created AI that allows to supress background noise perfectly, but Google with their billions is unable to even marginally improve their lackluster close-captions solution?", considering they are already snooping on your conversations constantly through your phone so it already is trained to identify human speech
The squeaky oil gets the wheel, as the going says. I'm sure there are many that appreciate this change, and appreciate your previous work.
Philip, you can make Whisper output directly to timed subtitles.
I just wish people weren't so lazy with AI subtitles. They're fine if you actually read over them and correct any mistakes but often people don't and it makes it very difficult to understand some things, especially when they're displayed as if a human wrote them and theres no indication that they're auto generated
As someone who uses subtitles often I do try to say when something is wrong in a nice way. Generally it is like "There is an error in the subtitles at X. It should say [Y] instead of [Z]." Instead of just "The subtitles are wrong." And if the subtitles show something completely different I will just say that because in those cases it was just missed during the subtitle process/something went very wrong. Like, one time I was watching a video and they seemed to have accidently copied their source into the subtitle so I just said it completely messed up.
Also, on the fact it is a thankless task, if I remember to I do like thanking people for adding subtitles because it takes a lot of effort and most people do not think it is worth it. Even though it is great because it allows people who regularily could not watch it to be able to….'cause RUclips's autogenerated subtitles _SUCK!_
Rambling. I think a lot about subtitles and RUclips. You do them pretty well, I do appreciate them even if I do not neccesarily need them.
As someone who used your subtitles for years, seeing it go away is saddening yet understandable. If you do use tools to help you subtitling though, I don't mind! Being not a native English speaker, I usually use subtitles as a supplement to my understanding of the narration. Sometimes words just slip through my mind and subtitles help me focus better, weirdly enough.
1:00 just as you said that I said it myself too. This is one of those huge areas AI frees people from working on. RUclips is weird with rolling out new features or ginormous changes to existing areas of function. It usually comes in waves every few years and likely after plenty of A-B user testing to garner feedback (or quietly via telemetry). Its a scary future but its not going to be 'all' bad.
I'm far more attracted to training my own models on data for local use than (As usual, thanks Internet) forfeiting my data to some one-click site with a nasty ToS. While this isn't always something that matters I've become more and more aware of these potential security risks and data leaks than ever these days.
That Thumbnail is disturbing as hell. Inspiration was The Ring, I guess?
And, although it might feel thankless, non-native speakers and all kinds of people are very thankful that you provide thorough subtitles. You just don't get to hear it that often, I guess.
So, thank you!
Thanks, asked for a prompt of 'Computer sucking the life force from books'
@@kliksphilipCreepy. Horror movies will be SO disturbing, soon.
1:30 "The subtitles written for this video" was the first mistake I noticed, but even before that, I noticed the overall formatting seems less YT friendly, like it's YT auto captioning masquerading as manual subtitles that some low effort channels do. There's not as many natural starts and stops with each sentence, which makes it harder to follow.
While text is fine in most cases, some manual correction is still required, ESPECIALLY for timing the captions.
Not sure how long it takes for you to generate the subtitles on your own PC now, but you could give fast-whisper a try. It builds atop whisper and I know that it is really great and generates subtitles in seconds instead of minutes.
There is now also insanely-fast-whisper, not sure how that is, but I think it's also very good 😄
I was watching and going "Yeah, it looks like these subtitles are pretty good. I'd expect more wrong if this was a fan translation/sub, but this seems adequate." then when I went to the comments, I realized I was watching RUclips auto-generated subs.
I've seen some horrible auto-generated subs, and I can imagine that this video might have just been an example of when RUclips can do it right. Hoping this process has improved things for when it doesn't work.
I will be honest with you Philip, the only reason I always watched your videos with subtitles ON, even though I never needed them, was to spot either discrepancies for fun or small Easter Eggs.
Honestly i'm surprized that some people discover whisper only now, it's been like several years since it released
Some new LLMs like Gemini 1.5 are multi modal (video/audio too!) so maybe it can write it even better with more context.
For me, RUclips's auto-generated subtitles are pretty good, and only makes a few mistakes where the audio itself isn't clear, so I'm surprised there are people who actively hate this feature. Unless they draw the line with RUclips making any mistake at all?
RUclips already does reasonable auto-generated subs, that even pop up with almost perfect timing.
Many times, the custom subtitles on videos have mistakes and have bad timing, so I almost always prefer the auto-generated subs.
As long as you don't *specifically* remove them like some annoying channels, I'm happy.
Yes, I know they sometimes have mistakes, but they are still better than custom subs with bad timing which still have mistakes.
agreed. RUclips does this for you, manually placed subs should be saved for actual quality.
I prefer the auto generated ones because the way each word pops up as it's being said makes it much easier to follow along for me. RUclips also has some annoying bugs with captions staying on screen for the wrong duration when watching sped up, except in the case of the auto generated ones because they update so frequently.
I might be wrong but I think Whisper might actually be able to output per-word timings for transcriptions too, which with some light scripting could be turned into a .vtt file that could be imported into RUclips instead of relying on their auto-timing?
Adobe Premiere has a purchasable adon for generating subtitles. You can also export subtitles with timestamps to a file and import them into RUclips.
My "Live Captions" software (From FUTO) made this caption when you talked about the 8600G: (Timestamp 1:49)
I've noticed so far, it's mostly to do with processor names, like it kept lopping a zero off the names like the rising 8600 g, because saying 8600 is interpreted As 860 and not 8,600.
Which one do you use for local deployment? How well does it run on your systems? Could a mini-PC that serves as a NAS device be good enough to run such AI? Like if one were to make a home-server for note-transcribing? I would love more info on your local ai stuff, or resources you used to inform your own decisions there.
not to mention, youtube does this really annoying thing with its auto subtitles where it censors any and all swear words and any words it *thinks* are swear words and replaces them with [__] UNLESS the creator manually turns that feature off on their channel, a setting which most creators don't even know is enabled, especially the ones that make unscripted content.
RUclips needs a 'overwatch' style subtitles editor where users can correct as they see fit. They have a hidden elo that works with others to ensure accuracy. System is doomed with bot accounts i guess overpowering real people but perhaps something youtubes AI could also judge and deem accurate or not.
I know subtitles are important but for the sub 1% of people who actually need them, they probably aren't even the people complaining. Your time is precious, the trolls time is not.
at work we use a programm called "Dragon Naturally Speaking" - its local on ur PC and works pretty well -> if ur looking for an alternative
as someone who is hard of hearing and relies on subtitles (though is not completely deaf), i don't agree that youtube automatic subtitles are "still as bad as they are"--it is frankly very rare that i see a mistake from it (or it's form a completely reasonable scenario like multiple people talking over each other, which i would be interested to see if there's any models out there that can split those up well). i also very much appreciate that the subtitles come out as individual words, versus the subtitles on this video which are not grouped into any obvious chunks (but that's youtube's own fault with its automatic system, so meh). in this video, it got "smarmy" wrong, which whisper did as well, and "lopping" and "transcribing". not perfect, but far far far from bad--i regularly use it over creator's existing captions, even. furthermore, it handled the "freakishly" mistake that many people in the comments have already pointed out, without changing the entire meaning of the sentence.
Transcribing audio recordings is exactly the type of simple but extremely dull and time consuming task that AI is perfect for because normal automation is not capable of doing it properly (though to be fair to RUclips we've come a long way since when they first introduced automatically generated subtitles).
My guess as to why RUclips isn't already using AI for automatically generated subtitles is because of how much it would increase video processing time which matters given how many hours of footage are uploaded to RUclips every day.
Edit: I just saw the error at 2:30. I guess that shows you can't trust AI to do this task on its own.
subtitling is pain especially when i prefer writing the subtitles (in davinci resolve though because youtubes subtitle system is cholera) since i get to add who says what and sometimes speech that isnt as audible
took me hours to subtitle a 20 minute video... probably half of the total editing time for it
I hate that most AI programs are only accessible online or are complicated to run on your own PC
This isn't one of those thankfully. Whisper large-v3 is available to download and it's exactly the same as what's online.
@@mar2ck_ are you sure? maybe the word "complicated" means something else to you than it does to me? because I can't seem to find any simple way to install this thing on my computer
I don't expect AI to be able to transcribe proper nouns that aren't common in real life. The name of a map, or a character, or if you covered other video games, the names of abilities and items and worlds.
I'm always endlessly thankful for youtubers who make real subtitles, even if they occasionaly happen to have couple of errors
As someone that majored in translation/transcribing just before the AI revolution of past years, I can only say one thing... fuck me I'm gonna be jobless by the end of this year...
Ive rarely done subtitles cause its such dull work but having this new ai black magic is wonderful. The only issue with me is my European accent™️ so it can have a very hard time deciding which words I'm using
The way you tube separates the parsing on these subtitles per word means the auto generated ones are still more useful to me as someone who uses captions. With the way it doesn't properly separate out full ideas I find them pretty hard to follow. I don't know if you can get it to generate the progressive caption style transcript but that would be much better than RUclips's automatic pacing feature. When I've generated captions, I've usually used adobe though as they have pretty good caption editing tools and can generate per word captions easily.
Also I thought this was a trans video with how the thumbnail was laid out, and i was like what's trans cribing, why are we putting the trans people in cribs?
Running AI shit localy is actually so based. Technology that "lives in the cloud" may be convenient but the thing about clouds is that they tend to change their shape over time into something unrecognizable from what they were before, or, even worse, a cloud can disperse entirely as if it was never there in the first place. Clouds are temporary, hardware is permanent*
the queue on huggingface boutta be huge lmao
oh my gawd kliksphilip admitted to using AIIAIAIIAIII!!!!!111!
i appreciate your effort! most youtubers woulnt have bothered at all
1:30 "So see what you think of" is missed
you could diff the script and the ai transcript and look for any out of place mistakes rather than improvised lines
0:45 "occasional SMAMI comment", but yeah otherwise very accurate
I think a real test would be to find out how loudly you can play Caboosing over narration before the AI stops being able to detect words well :P
The real question is how good is whisper with multiple or overlapping voices, would be a godsend for me 😂
I wanted to use DaVinci Resolve's tools for this, but it's only in the paid Studio version. So I improvised and tried the desktop app for CapCut and it works shockingly well. Even with my more unique accent and manner of speaking, it picked it up really well and I just had to make a few edits and then export the SRT file. I'd love to see a comparison between that and this tool + RUclips.
Pretty crazy.. A friend of mine used to do subtitles for Netflix for years. Almost a career. That job is pretty much extinct already.
Microsoft's transcribing on the video platform you get with 365 is good.
The subs at 2:30 turned "...the whisper transcribe feature online is freakishly good" into "...the whisper transcribed feature online is free. It's not really good...". So it's got a sense of humor at least, even when it fails.
Maybe I'm just the guy who doesn't wanna do drugs at the party and just sits in a corner the whole time.
The reason subtitles are a “thankless task” is because they’re a tablestakes expectation for accessibility. I don’t think that comes from a place of entitlement, some people simply can’t consume your videos without them. I think every video ought to have at least good-enough subtitles, which you’ve done a generally fine job at. Personally I’ve found your subtitles to be good enough, the occasional error is not any more frequent than it is on any other channel. I think the people complaining in your RUclips comments are engaging in bad faith… shocker I know
As someone who is trying to make it as a small time creator, subtitles are certainly one of the more time consuming things. Editing is already such a slog, it can take easily 5-10 times as long to edit something as it took to record it.
That said, Do not use auto generated subtitles. People who genuinely need subtitles for accessibility deserve to have the proper subtitles. It's lazy to hit the generate button and then just put up whatever comes out. Auto generate them, sure, but ABSOLUTELY DO NOT UNDER ANY CIRCUMSTANCES EVER just put the raw subtitles on your video. It makes it look like you don't give a shit when people are reading the subtitles and suddenly random unrelated trash comes up.
You know what AI cant do? Hide extra jokes in the subtitles. But well, most of the times Id miss them anyway since I dont watch your videos with my eyes glued to the script.
I'm willing to bet philip made the 2:30 incident on purpose
You can give whisper AI an "initial prompt" to help it. Maybe put your actual scripts in there, so it can handle words it doesn't really understand better.
The issue is that this transcribing just won't yield fully accurate results. With the way the technology is in its current state, it will make a few mistakes per video. It's not THAT big of a deal, as the message will still probably get across fine, but it's just overall gonna be slightly not fully accurate for a while, and will definitely need manual tweaking to some degree. Maybe in the future.
Everyone talking about 2:30 but nobody realized at 2:39 in the document, that part of the video is correctly transcribed... aand you got me to comment about it. damn you!
EDIT: just noticed it was an honest mistake, when you override everything at 2:58
But keep doing it automated, it's what I would do. And it's a perfect use for it
If the video doesn't have problems with subtitles, people won't say anything about them, and that is normal.
what an interesting way to write “transcribing”
(previous comment got sniped let's hope you can still see it in some comment review queue or something) also copy+pasting the script into the prompt field will probably help a bunch
Dis a banger. See you soon future 😊