OpenMai, Sam Outman, and Navia Stokes
« previous post | next post »
Among the many interesting takes on OpenAI's claimed solution to the Navier-Stokes Millennium Prize Problem, I've been most interested in Sabine Hossenfelder's discussion on YouTube., partly because of the content but mainly because of Google AI's weird transcription. Its first sentence has two odd errors:
Open eye claims to have solved the Navia Stokes Millennium problem.
Elsewhere in the transcript we get 'OpenMai', 'openi', and 'Omei' as references to OpenAI, which you'd think Gemini would "know" about. And Gemini somehow "knows" to capitalize "Navia Stokes Millennium", but not to spell Navier correctly.
There are plenty of other errors, some odder than others, e.g. 'Sam Outman' for "Sam Altman", as you can see for yourself in the transcript and subtitling.
And yes, I know that Sabine's English is accented, but that's exactly the kind of thing that a Language Model is meant to deal with. And here's a native-English-speaker's YouTube commentary on the Navier-Stokes solution (and the allegations that OpenAI stole aspects of the solution from researchers at NYU and Anthropic), in which OpenAI appears several times as 'Open AAI', and "Navier Stokes" is variously rendered as 'Navia Stokes', 'Nabi Stokes', and 'Nabier Stokes'.
Here's what Gemini says about the current automated transcription process:
Q: How does Google now create Youtube transcripts ?
-
- Audio Extraction: YouTube pulls the audio track from newly uploaded or existing videos.
-
- Speech Recognition: Google uses specialized speech-to-text AI models (evolving from older models like Chirp to advanced tools like Gemini 3.5 Transcribe) to listen to the spoken words.
-
- Timestamp Generation: The system matches the recognized words with specific points in the video's timeline to create clickable timestamps.
-
- Automatic Publishing: YouTube makes the auto-generated transcript available in a side panel for viewers to read, search, or copy. Creators can also edit or upload custom transcript files.
I could say more about why such embarrassing goofs are now common (basically the text tokenization side of end-to-end models, and their limited context window), but that's for another time.
Update — For the commenters who attributed the transcription failures to YouTube's choice to use a relatively cheap and low-quality STT system, here's some evidence that you're right, in the form of Gemini 3.5 Transcribe's output for the (audio channel of the) cited video, which lacks most of the cited errors, though it still makes a few mistakes. Google would charge me about $0.025 for an API-driven version of this effort, which I suspect Sabine would be willing to pay.
Michael Watts said,
September 10, 2026 @ 7:21 am
Well, I'd have to agree with the AI transcription that Sabine Hossenfelder's pronunciations of "OpenAI" are like nothing I would ever expect.
It's somewhat challenging to understand her in general, but that word in specific goes well beyond "somewhat challenging". I think the fact that the AI transcription correctly transcribes her as producing nonsense with obvious /m/ insertions reflects well on it, not poorly. That's what she's saying!
I already know what it's like to have input / transcription systems assume that the signal they're getting just can't be real, and therefore transcribe something entirely different. I don't want that. I like my errors to be obvious and easy to resolve.
This problem spans several levels. At one level we want a system that can turn a raw sound wave into phonemes. At another level we want a system that can turn /oʊpən/ into "open" and /eɪaɪ/ into "AI". We may or may not want a system that can turn /oʊpəneɪaɪ/ into "OpenAI"; the name feels like more of a word-by-coincidence. There will never be a system that can accurately transcribe arbitrary brand names it doesn't have preexisting knowledge of.
When the sound -> phonemes level gives us /oʊʔmeɪaɪ/, we don't want a system that will turn that into "OpenAI", or into "open AI". We might want a third level above the other two that can resolve processes like "well, 'OpenAI' would make sense in context here, and this string of sounds with no valid interpretation isn't thaaat far off"; this is basically my understanding of how a human would handle the task.
Michael Watts said,
September 10, 2026 @ 7:29 am
Actually, this reminds me of something I noticed about "Baden-Württemberg". I have no hope of producing an /m/ assimilated to the following /b/ in "Württemberg". I have to turn that into an /n/ assimilated to the /t/ preceding it. (And of course the /t/ must become a glottal stop since it's followed by /n/, but I think that happens in the German too?)
Paul Clapham said,
September 10, 2026 @ 9:57 am
That's nothing compared to say the AI transcription of the summaries online of the Vuelta a Espana cycle race. You are unlikely to recognize the various bastardized versions of "Wout Van Aert", for example, transcribed from the high-speed reportage.
Phillip Helbig said,
September 10, 2026 @ 10:27 am
https://www.youtube.com/watch?v=6nSKkwzwdW4
Phillip Helbig said,
September 10, 2026 @ 10:44 am
She does have a rather strong German accent. She also mentioned publicly that she has some sort of confenital defect which prevents her from opening her mouth very wide. The two together can be confusing.
Check out Indrainamerica which is a video blog about a German woman living in the States. I first came across her in some tongue-in-cheek but very accurate short videos comparing various customs in Germany and the States (and they are so in favour of the former that I wonder why she is in the latter). She plays the roles of a German woman and a U.S. woman. At first it wasn’t clear to me if she was an American doing a really convincing German accent or a German doing a really convincing American accent. It’s the latter. There is also a video in which how she explains how she managed that. (Note that the latter is more difficult, as a native accent is spoken much more and is objectively correct, whereas foreign accents vary depending on the speakers’ abilities and none is objectively correct.)
Phillip Helbig said,
September 10, 2026 @ 10:46 am
confenital —> congenital
Coby said,
September 10, 2026 @ 11:31 am
@Paul Clapham: in the American commentary on la Vuelta, one commentator's pronunciation of Enric Mas is transcribed correctly, but the other's comes across as Henrik Moss.
David Marjanović said,
September 10, 2026 @ 2:02 pm
Interesting. For me it's close to the opposite; when I actively try to put /n/ there, I'm more likely to produce [nm] than just [n] if I don't pay a lot of attention. But then, my native southeastern German merrily assimilates nasals across word boundaries, while farther northwest that doesn't even happen in compound nouns, apparently.
That seems to be a specifically Rhineland thing.* The other accents just do a nasal release on the plosive; and that's what Hossenfelder does in the video. All three diphthongs in OpenAI come out right, but the -pen- part is [pm̩].
* It's quite fascinating: -pen, -ten, -ken become [ʔm̩ ʔn̩ ʔŋ̩] there, i.e., the place of articulation is preserved by the syllabic nasal but disappears from the plosive. What is more, -den also becomes [ʔn̩], but possibly with a lenis instead of fortis [ʔ]. That would be an amazing phoneme contrast if it's real; I haven't heard enough to be sure it is.
Y said,
September 10, 2026 @ 4:14 pm
I hate to defend Google and YouTube, but surely the volume of videos needed to be processed requires a stripped-down text-to-speech model. In other words, they need to trade quality for the necessary quantity.
Michael Watts said,
September 10, 2026 @ 5:55 pm
I want to stick up for the concept of the non-phonemic contrast, a distinction that is maintained by speakers but nevertheless not phonemic.
The first example that comes to mind is "our" vs "are" in American English, where you have the option of applying lip rounding to "our" and don't have that option with "are". In my model, the two words are not phonemically distinct, but some speakers may distinguish them systematically in running speech and a larger group of speakers may distinguish them when intentionally trying to do so.
Another might be the words "thirteen" and "fourteen", which are sometimes described as being pronounced with geminated /t/ mid-word. I can see why they're described that way, but consonant gemination in general is not really an English phenomenon, and the case for a non-phonemic contrast is strengthened here by the fact that "thirteen" and "fourteen" are commonly confused (on the listener's end) with "thirty" and "forty".
Kenny Easwaran said,
September 10, 2026 @ 7:18 pm
I recently watched a video about Tesla's new Cybercab, and every time someone said "Cybercab", the captions wrote "Cybertruck", because apparently the latter is in the dictionary but the former is not!
AntC said,
September 11, 2026 @ 12:50 am
She [Sabine] does have a rather strong German accent. She also mentioned publicly that she has some sort of confenital defect which prevents her from opening her mouth very wide. The two together can be confusing.
It goes a lot further than that. Sabine seems not to understand that English has phonemic stress. So for example she stresses 'colleague' on the second syllable. There was a whole video about Arsenic in which she stressed that on the second syllable (as in German) — which I at first took to be about to say 'arsenical' then had to rewind and re-parse; by which time I'd lost what the sentence was saying.
You'd'a thought she'd employ someone in her editorial team with good command of English; they're surely not hard to find in Germany(?)
I can in general cope with foreign accents (after all English doesn't have a leg to stand on when it comes to perplexing locutions); some of my best friends have strong German accents; but the combination of plain weird diction with subject-matter on the edge of my knowledge is often just too hard.
VVOV said,
September 11, 2026 @ 5:32 am
Do German speakers, when speaking in German, generally pronounce “OpenAI” like Sabine does? She clearly is someone who talks about OpenAI frequently, and her [əʊʔmeaɪ] sounds much farther “off” from the English pronunciation than the rest of her English generally does to me.
I didn’t otherwise find the speech in the video difficult to understand. But I’d point out that part of why it sounds odd is not only her accent or word-level stress, but also slightly unnatural prosody that i think arises from *reading a pre-prepared script off a screen*. As almost always in videos that primarily consist of a person reading to the camera, it makes me wish i were consuming the same content at my own leisure in the form of a transcript or printed essay.
bks said,
September 11, 2026 @ 5:35 am
Maybe it's just me (native English speaker) but I find it very hard to say "ChatGPT" in fluid speech.
Tom said,
September 11, 2026 @ 11:18 am
@Kenny Easwaran-
Yes! I have noticed the same phenomenon in other videos. It's very interesting. I know next to nothing about the tech behind these transcriptions, but it is clearly not working the way a person would intuitively assume.
Another thing I have recognized is that recent auto-generated video transcriptions in YouTube compared to older ones seem to trade decreased accuracy in proper nouns for increased accuracy in punctuation that models the natural speech. Not sure why this trade-off happens.
David Marjanović said,
September 11, 2026 @ 12:01 pm
German, too, has phonemic stress – and Kollege is stressed on the second syllable. This is reinforced by the long-looking English spelling with ea. English stress in Latinate words is very often hard to guess from German (or French or actual Latin).
Sure.
I insist: not -[ʔme]-, but -[pm̩eɪ̯]- – labial plosive, syllabic [m], FACE diphthong.
ohwilleke said,
September 11, 2026 @ 2:24 pm
FWIW, as an attorney who regularly works with human court reporters transcribing spoken statements, getting proper names transcribed and spelled correctly is the most common challenge and source of mistakes by far in transcriptions, so it isn't surprising than an automated system to transcribe speech would experience the same challenges.
Michael Watts said,
September 12, 2026 @ 4:22 am
I think I'm with you on the syllabic [m] and I have no real opinion on the quality of the diphthong. It's not hard to understand that she means to be producing FACE.
But I don't think it makes sense to quibble over the entry to an unreleased stop. She's speaking English and all unreleased stops are /ʔ/. Asking whether the glottal stop is "really" the first half of a /p/ or a /t/ or a /k/ is like asking whether the middle consonant in /pɛɾl̩/ is "really" a /t/ or /d/: there's a correct answer, and the correct answer is "no".
One easy example of this is that I often pronounce "welp" (the informal pronunciation of the discourse particle "well") with open lips, which given the articulation required for /l/ might be better spelled "welt". No difference. It's /wɛlʔ/ no matter what your lips are doing.
And I would also say that the reason you're seeing us transcribe her as saying /oʊʔme(ɪ)aɪ/ is that English phonology just doesn't allow for /oʊʔm̩/. I'm willing to agree that that's what she's saying, but that kind of thing is precisely why we're identifying her as hard to understand.
Andrew Usher said,
September 12, 2026 @ 10:19 pm
First, the least important point: She uses, it seems, both monophthongal and diphthonal FACE, which is fine, but when it comes directly before another vowel as in 'AI', the diphthongal variant is really needed to provide the linking glide, there mandatory (unless a glottal stop is used) in even slightly careful speech.
And yes, that's a syllabic /m/ in 'open', also un-English as we have no choice but /opən/ or the like. If it were opem the syllabic consonant would, I think, be possible but still not usual. Spoken German may use syllabic nasals wherever possible but again English does not; this may account for the difficult Michael Watts reported with "Württemberg"; I agree if one tries to force the /m/ to be syllabic, but adopting the more natural (to English phonology) alternative with a schwa removes the difficulty and I easily say /vʏrtəmbɛrk/ (the rhotics may be taps or approximants) with no intention of sounding like a real German speaker.
I do not understand his above comments that unreleased stops are really /ʔ/. There wouldn't by a separate term for them if they were, I think, and I can attest definitely that e.g. 'pep', 'pet', 'peck' sound different even though the stops are all unreleased (and pre-glottalised too!) even if it's only the effect on the preceding vowel that is heard. Voiced stops can equally be unreleased and are differentiated by that and their (partial) voicing. For what it's worth, I hear in her 'OpenAI' /p/ and not /ʔ/, though I'm not completely certain without a spectrogram.
To the general issue, I don't think automated transcription systems are appropriately criticised for how they work. They aren't being used for the process but for the final product – here, what we want is something at least comparable to a decent human transcription. And most definitely that would not get a familiar proper name like 'OpenAI' wrong, no matter the accent, as long as it was recognisable, which it is. 'Navier-Stokes' and 'Sam Altman' would also be checked (if not known already) by a good human transcriber, and in general getting knowable proper names right should be one of the less difficult parts of automated transcription. But in the current AI-hype world, LLMs and unlimited cycles are thought to be the answer to everything, and I'd guess that to be the most fundamental problem here.
David Marjanović said,
September 13, 2026 @ 12:36 pm
My point is it's not unreleased – it has a nasal release. That doesn't occur in L1 English; and that, in turn, is a fact few L2 learners manage to guess. (We're of course taught that kn- is just /n/, but we're not told that that's part of a broader phenomenon.)
Andrew Usher said,
September 13, 2026 @ 8:27 pm
I have no complaint about her pronunciation being called nasal release, but if you mean that nasal release doesn't occur in L1 English, it isn't so – it's just rarely possible when it would cause assimilation.
Into a syllabic nasal it's a possibility in /tn/ and /dn/ and often the prescribed option in words of the type sudden, Sutton. Into a non-syllabic nasal it can naturally occur only across a syllable boundary but as far as I know is unrestricted there – such clusters need not be broken by /ʔ/ or anything else; and even if most people don't manage it there's nothing either wrong or odd about saying technical with [kn].
So I'm not sure that the loss of the stop in initial /kn/, /gn/ really _is_ "part of a broader phenomenon" rather than affecting only those clusters.
Jarek Weckwerth said,
September 15, 2026 @ 9:53 am
Many standard descriptions of English phonetics I'm aware of (e.g. Cruttenden, Roach, Wells) allow syllabic m at least from in happen, Cobham and similar words.
And of course @Andrew Usher annd @David Marjanović are 100% correct in saying that unreleased plosives are different from glottal stops. (But, informally and anecdotally, the conflation of the two seems to be sometimes found in American linguists. — I'm being nasty here…)
Lack of release doesn't need glottalization, as the lenis series demonstrates. The perception of place comes from the articulatory movement in the vowel towards that place and doesn't need release. A glottal stop may but doesn't have to be accompanied by oral closure, but if it isn't, then in English it's very typically perceived as a /t/.
BTW I remember an informal conversation with the late Ian Maddieson in which he pointed out, with his tongue firmly in his cheek, that "unreleased stop" is not a particularly fortunate term. Not releasing an oral stop leads to death from suffocation ;) "Inaudibly released" is a better term.
Andrew Usher said,
September 15, 2026 @ 7:48 pm
Jarek Weckwerth writes:
I think this thread demonstrates that syllabic /m/ in 'open' (which has the same structure as 'happen') is considered non-English; admittedly in 'OpenAI' it's immediately followed by another vowel, and would be less conspicuous before a pause or a consonant.
As I have already noted with regard to a hypothetical opem, when already homorganic as in 'Cobham', it would be less marked but still seem unusual.
Michael Watts said,
September 16, 2026 @ 12:41 am
They are not different by the time you're working with phonemes. As such, you cannot sensibly expect an American listener hearing English to draw any such distinction, nor would any process that produces a sequence of phonemes be able to do so.
Indeed, nasal release of /d/ is the norm. You could insert a vowel, but you'd sound strange.
In my variety of AmE (and this goes well beyond me personally), a /tn̩/ sequence obligatorily reduces the /t/ to a glottal stop. There is no alternative in natural speech. You could produce e.g. "mountain" with an aspirated /t/ followed by a vowel, but it would be clear that you're overenunciating for some reason that is hopefully obvious in context. (My personal instinct is that this would sound less strange, but be more wrong, than pronouncing hidden without nasal release.)
For technical you have more freedom; you're free to produce [kʰn] across the syllable break or [ʔn] as you see fit.
If you're willing to run an experiment, you can try using words analogous to "tetnical" /teʔnɪkl̩/, where what is formally a reduced /k/ is pronounced as a reduced /t/ instead, in conversation with native speakers and see how confusing it is. I would expect no loss of understanding, with reactions varying from "minor bemusement" to "didn't notice".
I mean, neither of those things is true of nasal release. You can hear it just fine. And I see no problem with the concept of an "unreleased stop" which is then followed by the cessation of speech and a return to normal breathing.
But that's a very fine and ultimately pointless line to draw. We can just accept that when we talk about "release", we're only referring to releases that have some effect on the sound.
(As an even finer technicality, you can breathe just fine without releasing /t/, and that fact is what makes lateral sounds possible.)
I can say that I can produce /hæʔm/ with syllabic /m/, but I never would; if I heard it, I would probably understand it correctly but it would mark the speaker as having a foreign accent. I was able to produce /kɑbm/ with syllabic /m/, but that took me several tries; it is not at all a natural sequence of sounds for me. I don't think nasal release of /b/ into syllabic /m/ is really allowed. (Into non-syllabic /m/, something like "Robmeister", it's fine.)
Daniel Barkalow said,
September 17, 2026 @ 11:38 am
The aspect that I find strange about AI transcripts is that they don't pick up the cues that a large number of points in the transcript are talking about the same thing and use this to collect multiple instances of the same speaker saying the same thing, vote on the best match among them, and use that consistently. Surely the second video isn't talking at different times separately about all of 'Navia Stokes', 'Nabi Stokes', and 'Nabier Stokes', but rather continuously about something that might have been any of these, but is most likely the best fit for the time it is said most emphatically.
Andrew Usher said,
September 17, 2026 @ 7:46 pm
That's probably a harder problem than you think, requiring not only a video-long context window but knowledge of linguistic structure that would allow determine that all those instances of Navier-Stokes likely refer to the same thing despite being in all different sorts of grammatical situations.
What would be reasonable would be to add words/names identified in the video to the working dictionary, and perhaps even correct the entire transcript based on that. That should achieve the desired result in most cases, that at least the spelling would be consistent. However it might be even easier to abandon the dogmatic AI-only approach and manually add 'Navier-Stokes', either in general or for particular videos that mention it.
Jarek Weckwerth said,
September 21, 2026 @ 2:11 pm
@Michael Watts
They are not different by the time you're working with phonemes. — Obviously, because we're working with allophones here.
The /t/ in mountain and /k/ in technical would not be aspirated in the usual sense. They would be released but release is not tantamount to aspiration, as is shown by lenis plosives.
followed by the cessation of speech and a return to normal breathing — Well, the "tongue in cheek" reference in my original post meant that Maddieson was not speaking fully seriously.
you can breathe just fine without releasing /t/, and that fact is what makes lateral sounds possible — You're thinking of the tip of the tongue only. If there is lateral opening, you're not making a [t] but a [l] or a voiceless [l]. In a [t], there is bracing, i.e. the sides of the tongue are in full contact with the sides of the palate, and there is velic closure. All of this prevents the escape of any air, and therefore breathing, and enables the buildup of pressure behind this complete seal, i.e. the first part of a plosive consonant. (You've just done an introductory phonetics exercise for discovering the proprioceptive aspects of the articulation of laterals, btw.)
Here is an example of an American English speaker, from the first few hits on Youglish, doing a [bm] in happened by.
Jarek Weckwerth said,
September 21, 2026 @ 2:24 pm
Of course I meant [pm].
And here an excellent example of open book from a British speaker where you have a very clear view of the lips. Would you say there's a vowel between the [p] and [m]?
Out of the first ten "UK" samples of open book at least two are evident [pm] and another three may arguably have a very minimal vowel (and thus no real nasal release) but a [m]. (Of course the second speaker is misidentified since she's a New Zealander).
Jarek Weckwerth said,
September 21, 2026 @ 2:37 pm
(Promise this is the last post — I should have thought and planned before spitting these out. Sorry.)
It's useful to remember that you can also have nasal release from a velar into a [ŋ], as demonstrated by can go here.
Michael Watts said,
October 5, 2026 @ 6:03 pm
Well…
There's no theoretical problem with such a release in a language where the sound sequence is possible.
It's certainly not possible in English, which doesn't allow for syllables to begin with [ŋ] under any circumstances.
And I don't see how you think "can go" could be an example of nasal release from a velar into a nasal velar. The nasal precedes the velar. (Yes, I watched the video. The pronunciation is just as you'd expect.)
Jarek Weckwerth said,
October 7, 2026 @ 1:08 pm
I hope you'll agree that /k/ is also velar. It's released nasally into the /ŋ/.