OpenAI's Whisper is another case study in Colonisation | Hacker News Reader

OpenAI's Whisper is another case study in Colonisation

blog.papareo.nz

17 pointstotetsu21 comments

Mozillas Common Voice Project (https://commonvoice.mozilla.org/en/languages) is creating an open dataset for many minority languages to make it easier to support them in STT systems without legal issues. If you speak one of these languages please consider donating a few minutes of your voice.

The article has some good points about marginalised languages and dangers of incorrect use, but it is wrapped in awful nationalist and tribal rhetoric. It divides the world into in/out groups based on heritage and claims total ownership of a language. This goes beyond protecting their heritage and promoting it, they want to keep it for themselves, never letting anyone else use it: "If anyone is to profit from te reo Māori it should be Māori and Māori alone". They decry developed societies making technological progress under a racial guise: "... open sourcing ... best serves those who already have a foot in the industry, which is mostly White and Asian men".

Claiming that you alone have a right to work with your native language is about as ridiculous as claiming that only Italians can make pasta. I don't see the point anyhow; who cares about the motivations of corporations if the product is good?

I don't know how it is with Maori, but I have seen many lament that for Sami (it's Sami people's day, by the way! Lihkku sámi álbmotbeivviin!), people are afraid to speak it because they don't have enough experience and are afraid to do it wrong.

It's not an entirely unreasonable concern that a language should get influenced/corrupted by a bad machine translator that is just good enough that people can pick up its errors. But humans can easily pick up errors from poor speakers as well... I think that as long as the prestige resides with experienced human speakers, the risk isn't great.

An interesting development, that. After American English or Scouse or Scots, all "corruptions" of standard British English that became established enough to be their own thing, we could soon get Machine Translation English. The Guardian will be very unhappy about it.

We developed a tool that lets you kind of understand other people if the audio quality is sufficient. With further training we can openly communicate with each other without fear of misunderstanding.

Obviously it has to be racist...

It's not unexpected that it is biased to the language of its creators. I'd expect similar results from Chinese AI, etc.

The question is it reasonable for English speakers with a generally European heritage to be the stewards of niche world cultures?

That seems like quite the burden, which is the opposite of strength.

>The question is it reasonable for English speakers with a generally European heritage to be the stewards of niche world cultures?

Something better than nothing? It's very unlikely that a tribe in Africa or Polynesia is going to create their own ML model anytime soon.

Although if you read ML papers, it's usually pretty diverse. The article is hinting at colonialism, especially as experienced by the Maori, i.e. largely "English speakers with West European/British heritage, also male". However, the original Whisper paper has 6 authors out of which 2 are East Asian, 2 are of Jewish heritage, 1 is a native Russian speaker, 1 woman. I'm not sure how they're related to colonialism in NZ.

The article reads like the author of the article has what's called "gatekeeper syndrome" in my country, i.e. they're just upset someone trains ML models on Maori language without their official approval/oversight because the author feels their foundation is entitled to speak for all the Maori.

I wouldn't consider the ai models to be the stewards of the English language. The people who speak English are. When you hear a bot speak in broken English, you recognize that it's broken English.

The stewards on the language will similarly hear the bot as broken, and people who don't know the language won't know the difference.

This article has an interesting premise (mindless data collection by big tech companies leading to very incorrect recreations of still existing cultures and languages "for show"/themeparking) but about midway through it becomes a bit of a mess in order to complain about certain pieces of tech/demands for the Maori to control every tech related aspect to innovation that could so much as touch on their language, that don't feel like they have as much substance.

Granted, looking at the website itself, this is likely because the article is chiefly meant to promote the organization the website belongs to, which is related to Maori language/culture preservation.

One interesting direction I think this article could've gone in is to focus more on what these sorts of mindless recreations can look like; the Scottish Wikipedia was largely edited by one American teenager and it resulted in a wiki that resembled "English ran through what a foreigner thinks is Scottish" more than actual Scottish. That's a real harm that only occurred because the Scottish Wikipedia doesn't seem to have been administered with a lot of attention given to it (because Wikipedia loves launching projects in all languages without really doing the legwork to establish a base for it). That's something that you can use to underpin why it's so important to have native speakers in these sorts of processes rather than jumping to demands that don't feel like they're as well build as the initial complaint against Whisper.

Some colonizer needs to teach the person who wrote this the importance of stating a clear thesis and backing it up with facts.

Halfway through the article it still seemed like a random hodge-podge of disconnected anecdata about a niche language and ethnic group.

The quoted student, who rejects consent of the copied artists and is paralleled to the colonialists of old (in the header image), is... Nigerian.

You'd think the race as politics people would get tired of generating these strange contradictions that they must formulate increasingly elaborate explanations for, but it's a fun little evergreen enterprise I suppose.

> we must respect data.

The author is a good 20 years late to the party I'm afraid

It baffles me why everyone nowadays thinks that whatever OpenAI releases is the first and groundbreaking. The fact is, Whisper is undesirable if you want to deploy it for business.

> Whisper is undesirable if you want to deploy it for business.

Why is that?

I find this an appalling take.

- You criticize making this transcription+translation model available on the grounds that this specific native language is so holy that they shouldn’t be permitted?!

- You deride in a lengthy way the model, which is powerful and more versatile than other state-of-the art models?!

- You get hung up on ridiculous details, like how the Maori language was not tagged as Māori in the code, and you keep using the Maori term for the Maori language?! (This is as ridiculous as insisting on calling every French cheese fromage, because the French do so.)

- You keep bringing up racist, colonial imageries, and it is just a harmfree language model?! It is not even a generative model. It is a transcription + translation model.

- You talk about being anchored in the community etc and gaining the trust of the indigenous etc. Why are you so holier-than-thou?!

- This is really long, without much content. You don’t make differentiated points. You’re just offended as a default setting and go out of your way to justify being offended.

The article reads like the authors think the language is dead and should only exist in a museum. A living and thriving language is going to change over time and develop dialects and accents. "Our model describes the one true language, and anyone who doesn't pass our test is wrong" does not allow for that

Whisper is not a mature enough to be criticized at this level. They are trying to solve a very hard problem, give them constructive feedback.

IPhone didn't support 100s of languages in the first few years.

The author is either entitled or resentful; neither helps.

So do they want AI to have knowledge of that language or not? The article is incoherent. It's clear the authors are mad for some reason, but that reason is not explicitly denoted in the article.

TLDR; Network effects apply to languages making it valuable to know a common language

Reply on news.ycombinator.com