The same happens with whisper-large-v3 on Chinese transcription: silence is transcribed to something like "please upvote, share and favourite this video". I suspect they trained the model on some random YouTube video without carefully picking really useful data.
In Chinese, it always added something like "For study/research purpose only. Please delete after 48 hours." This is what those volunteers added in subtitles of (pirated) movies/shows.
That is not the case here - I never encountered this with whisper-large-v3 or similar ASR models. Part of the reason, I guess, is that those subs are burnt into the movie, which makes them hard to extract. Standalone subs need the corresponding video resource to match the audio and text. So nothing is better than YouTube videos which are already aligned.
Indeed, with another model I would get persistent transcriptions of silent parts into 'Thanks for watching!' or '[MUSIC]'. Pretty dumb that this failure mode wasn't caught in some QA process, and there are now multiple transcription models suffering from the same issue. Having silent parts in your input audio seems like it should be a very common occurrence...
When I was taught mathematics, the zero value was always considered the most important edge case. You prove something for N=0 (or N=1), then for N=M+1.
It's even more important in audio DSP: processing near-zeroes can end up being extremely CPU intensive, look up denormal/subnormal floats.
When YouTube began building automatic transcriptions for captions, it regularly flagged any noise or music -- typically industrial noise -- with "[foreign]"
If it couldn't understand it, it was "foreign" for the longest time.
That's interesting, the few times I tried playing with whisper, I had the impression that YouTube style videos or random cellphone videos was something it did particularly bad with (compared to movies). My guess at the time was that most of the training material might be sub titles and raw screen plays.
The videos I tried to transcribe were also Mandarin Chinese, using whisper-large-v3. Besides the usual complaints that it would phonetically "mishear" things and generate nonsense, it was still surprisingly good, compared to other software I played around with.
That said, it would often invent names for the speakers and prefix their lines, or randomly switch between simplified and traditional Chinese. For the videos I tested, intermittent silence would often result in repeating the last line several times, or occasionally, it would insert direction cues (in English for some reason). I've never seen credits or anything like that.
In one video I transcribed, somebody had a cold and was sniffling. Whisper decided the person was crying (transcribed as "* crying *", a cough was turned into "* door closing *"). It then transcribed the next line as something quite unfriendly. It didn't do that anymore after I cut the sniffling out (but then the output switched back to traditional Chinese again).
Similar in the English model. Pretty clear they trained on YouTube videos where creators will put that in otherwise silent sections to ensure it shows up for people with CC on.
This reminds me, some years ago as Google was expanding its translation service, someone tried translating text into and out of an obscure African language (don't recall which) and it always came out as weird Biblical-sounding semi-gibberish.
My revelation was that machine translation needs a corpus of bilingual documents to learn from, and if the language is sufficiently obscure, there may not be any bilingual documents except for the Bible, which missionaries have translated into just about every language on Earth.
This is totally happening with other models too, at least with Spanish. Many transcriptions will end with something that roughly translates to "Thanks for watching!" even if it's never present in the original audio.
As I didn't see one correct definition of overfitting:
overfitting means that the model is too closely aligned to the test data, picked up noise and does not generalize well to *new, unseen* data. think students that learn to reproduce questions and their answers for a test instead of learning concepts and to transfer knowledge to new questions that include the same concepts.
while this sounds like overfitting, I'd just say it's garbage in, garbage out; wrong classification. the training data is shit and didn't have (enough) correct examples to learn from.
And it seems to be because the training data is largely unofficial subtitles from movies. Which often have a string like "Translated by X" at the end of the movie which is often silent while credits roll.
Looks like they used more official sources for German - there, silence is apparently hallucinated as "Untertitelung des ZDF für funk, 2017" according to one of the comments on the issue. Which makes sense, as the public broadcasters' "Mediathek" is probably the largest freely available resource of subtitled videos in Germany. I wonder if the ZDF gave its approval for it being used for LLM training though?
Let's not forget that some of real pirates (for example, corsairs) also were legal and performed legitimate pirate activities to ships of foreign countries.
How is this evidence of that fact? Honest question.
I can see how this might show that subtitles from online sub communities are used, or that maybe even original subtitles from e.g. DVDs are used. But isn't it already known and admitted (and allowed?) that AI uses all sorts of copyrighted material to train models?
The Chinese subtitles for silence use a common mark for pirated media in that language, according to other commentors here. In general it's pretty likely that if you're finding non professional subtitles they were distributed with pirated media in some form, that's where you get the most fan subs after all
The contention is that the specific translated text appears largely from illegal translations (i.e., fansubs) and not from authorized translations. And from a legal perspective, that would basically mean there's no way they could legally have appropriated that material.
> But isn't it already known and admitted (and allowed?) that AI uses all sorts of copyrighted material to train models?
Technically, everything is copyrighted. But your question is really about permission. Some of the known corpuses for AI training include known pirate materials (e.g., libgen), but it's not known whether or not the AI companies are filtering out those materials from training. There's a large clutch of cases ongoing right now about whether or not AI training is fair use or not, and the ones that have resolved at this point have done so on technical grounds rather than answering the question at stake.
Whisper is unusable IMO because of the hallucinations. Widely documented. Removing silence from audio clips helps, but even then it will auto correct grammar, translating bilingual speech, etc. Improved in the latest audio models but not solved [1]
I wouldn't describe it as "unusable" so much as needing to understand its constraints and how to work around them. I built a business on top of Whisper [1] and one of the early key insights was to implement a good voice activity detection (VAD) model in order to reduce Whisper's hallucinations on silence.
That's the problem with raws large models, it should always be coupled with satellite small models and logic. It's (probably) easier to detect hallucinations using a traditional ML/DL model that can catch mismatches (it's easy to build a synthetic dataset for this) than transcribing. And the simplest piece of code can detect a silence and that it should match no text.
In Russian it often hallucinates "Субтитры сделал DimaTorzok" ("Subtitles by DimaTorzok") at the end of things. Interestingly, I wasn't able to find any YouTube videos with that name in the subtitles, so it's not like it's in a lot of training data.
I tried googling this and found questions from Telegram users why voice messages recognition sometimes produces this phrase and who is this person. Also I found this thread [1] claiming that the subtitles by DimaTorzok are coming from some Russian youtube videos on gaming like [2].
Readit News
It's even more important in audio DSP: processing near-zeroes can end up being extremely CPU intensive, look up denormal/subnormal floats.
If it couldn't understand it, it was "foreign" for the longest time.
The videos I tried to transcribe were also Mandarin Chinese, using whisper-large-v3. Besides the usual complaints that it would phonetically "mishear" things and generate nonsense, it was still surprisingly good, compared to other software I played around with.
That said, it would often invent names for the speakers and prefix their lines, or randomly switch between simplified and traditional Chinese. For the videos I tested, intermittent silence would often result in repeating the last line several times, or occasionally, it would insert direction cues (in English for some reason). I've never seen credits or anything like that.
In one video I transcribed, somebody had a cold and was sniffling. Whisper decided the person was crying (transcribed as "* crying *", a cough was turned into "* door closing *"). It then transcribed the next line as something quite unfriendly. It didn't do that anymore after I cut the sniffling out (but then the output switched back to traditional Chinese again).
They trained the model on every YouTube video they could, and hoped the aggregate was useful data.
My revelation was that machine translation needs a corpus of bilingual documents to learn from, and if the language is sufficiently obscure, there may not be any bilingual documents except for the Bible, which missionaries have translated into just about every language on Earth.
Deleted Comment
It's the LLM equivalent of thinking that an out-of-office reply is the translation: https://www.theguardian.com/theguardian/2008/nov/01/5
Instead, it reverted to what it has seen before (in the training data), hence the overfit.
"Translated by Nancy Qanfar"
overfitting means that the model is too closely aligned to the test data, picked up noise and does not generalize well to *new, unseen* data. think students that learn to reproduce questions and their answers for a test instead of learning concepts and to transfer knowledge to new questions that include the same concepts.
while this sounds like overfitting, I'd just say it's garbage in, garbage out; wrong classification. the training data is shit and didn't have (enough) correct examples to learn from.
The Arabic text "رجمة نانسي قنقر" translates to English as: "Nancy Qanqar's translation" or "Translation by Nancy Qanqar"
"رجمة" means "translation" and "نانسي قنقر" is the name "Nancy Qanqar"
The MPA must be so proud.
Deleted Comment
Legally, why wouldn't they be able to do the piracy parts in one of those jurisdictions and then ship the outputs back to the mothership?
I can see how this might show that subtitles from online sub communities are used, or that maybe even original subtitles from e.g. DVDs are used. But isn't it already known and admitted (and allowed?) that AI uses all sorts of copyrighted material to train models?
Indeed, the captioning is copyrighted work and you are not legally allowed to copy and redistribute it.
> But isn't it already known and admitted (and allowed?)
No, and I don't see where you got that from. Meta [1], OpenAI [2] and everybody else is being sued as we speak.
1: https://petapixel.com/2025/01/10/lawsuit-alleges-mark-zucker...
2: https://www.reuters.com/legal/litigation/openai-hit-with-new...
The contention is that the specific translated text appears largely from illegal translations (i.e., fansubs) and not from authorized translations. And from a legal perspective, that would basically mean there's no way they could legally have appropriated that material.
> But isn't it already known and admitted (and allowed?) that AI uses all sorts of copyrighted material to train models?
Technically, everything is copyrighted. But your question is really about permission. Some of the known corpuses for AI training include known pirate materials (e.g., libgen), but it's not known whether or not the AI companies are filtering out those materials from training. There's a large clutch of cases ongoing right now about whether or not AI training is fair use or not, and the ones that have resolved at this point have done so on technical grounds rather than answering the question at stake.
Dead Comment
1. https://news.ycombinator.com/item?id=43427376
[1] https://speechischeap.com
Way to go Nancy! Keep up the good work, ya crazy bastard!
"[ sub by sk cn2 ]"
or
"Anyways, thanks for watching! Please subscribe and like! Thanks for watching! Bye!"
or
"This is the end of the video. Thank you for watching. If you enjoyed this video, please subscribe to the channel. Thank you."
[1] https://github.com/openai/whisper/discussions/2372
[2] https://www.youtube.com/watch?v=FAqyUuahMlc&t=401s