Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

github.com

Post reply on HN
Print view View original

81–90 of 232 posts

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#81

Post by ccooffee »

egberts1 wrote: post #67
crazygringo wrote: post #62

Earlier quoted context omitted.

Perhaps you could help by inventing a written version of ASL, since one doesn't currently exist. [EDIT: I was wrong, sorry, see my response below.] Seems like that would be a prerequisite for a model based on written language. And of course if you could also create an entire corpus of training material in your written ASL?

It is called SignWriting for all you multiple naysayers. https://en.m.wikipedia.org/wiki/SignWriting

Very fair point. I've never seen it used, but SignWriting is already in the Unicode standard[0] (at U+1D800 - U+1DAAF).

I suspect Meta/Facebook doesn't have a lot of content to work off of. I've only been able to find community-generated examples of SignWriting on the official website[1], and none of those seem to be using Unicode characters. MMS is an audio-to-text tool, so it seems unlikely that it can be trivially expanded to take in visual data (pictures of text or video of ASL being performed).

I suspect the process of turning viewed ASL into SignWriting text will be very difficult to automate. I would not be surprised if such a project would either use a different textual encoding or directly translate out to English (which also sounds terribly hard, but these LLM advances recently have surprised me).

[0] https://www.unicode.org/charts/PDF/U1D800.pdf

[1] https://signwriting.org/#RecentPostingsUS

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#82

Post by MauranKilom »

mach1ne wrote: post #23
generalizations wrote: post #6

Earlier quoted context omitted.

Based on the availability of STT, TTS, and translation models available to download in that github repo, a real-life babelfish is 'only' some glue code away. Wild times we're living in.

Nah, full-on babelfish is simply not possible. The meaning of the beginning of a sentence can be modified retroactively by the end of the sentence. This means that the Babelfish must either be greatly delayed or awkwardly correct itself every once in a while.

And sometimes that modification is not just "we don't know which verb it ends in" but "the whole structure is different than expected":

https://en.wikipedia.org/wiki/Garden-path_sentence

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#83

Post by tmpz22 »

crazygringo wrote: post #56
m3kw9 wrote: post #36

The problem with all these model releases is they have no demos or even video of it working. It’s all just download it and run it, like it’s an app.

They're intended for researchers/professionals not consumers, and I'm not sure how a video is going to be helpful? And the issue with a live demo is that these are resource-intensive, they're not just webpages. It's an entire project to figure out how to host them, scale them to handle peaks, pay for them, implement rate-limiting, and so forth. For the intended audience, download-and-run-it doesn't seem like an issue…

Its so weird to me that they'd do 99% of the effort and just skip past the 1% of work to provide a dumbed-down summary and instructions for broader appeal. Clearly these are released in part for public-relations and industry clout.

Don't get me wrong they published this, it took a ton of work, they didn't have to do it. But its ultimately a form of gatekeeping that seems to come straight out of Academia. And honestly, that part of Academia sucks.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#84

Post by reaperman »

EMIRELADERO wrote: post #21
reaperman wrote: post #9

Earlier quoted context omitted.

Id imagine you could use inference from this as training for a commercial model, as that isn't currently protected under copyright.

I'd go a step further and say that the models themselves probably aren't copyrightable.

I mean I personally feel like almost all software is math and anything that isn’t frontend isn’t legally copyrightable or patentable. But the courts disagree with my interpretation. And if it were up to me all copyright and patents would have exponentially increasing fees on an annual basis to maintain, but thatd require a change is statutes. I think the jury is still out on your interpretation though, and I suspect the courts will fall in line with whatever the largest rights holders want.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#85

Post by yamazakiwi »

egberts1 wrote: post #67
crazygringo wrote: post #62

Earlier quoted context omitted.

Perhaps you could help by inventing a written version of ASL, since one doesn't currently exist. [EDIT: I was wrong, sorry, see my response below.] Seems like that would be a prerequisite for a model based on written language. And of course if you could also create an entire corpus of training material in your written ASL?

It is called SignWriting for all you multiple naysayers. https://en.m.wikipedia.org/wiki/SignWriting

Most hearing impaired people have never heard or don't care to use SignWriting. You are right about it's existence, for what it's worth.

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#86

Post by fulafel »

echelon wrote: post #4

> The MMS code and model weights are released under the CC-BY-NC 4.0 license. Huge bummer. Prevents almost everyone from using this and recouping their costs. I suppose motivated teams could reproduce the paper in a clean room, but that might also be subject to patents.

My layman's reading is that the license does seem to allow recouping of costs, would be interested if there is a more nuanced interpretation to the contrary.

(tangentially, like another comment briefly mentioned, are models actually copyrightable? Programs are, because they are human authored creative works.)

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#89

Post by mkl »

neom wrote: post #7
cleverwebble wrote: post #3

Wow, I didn't even know there was 7,000 documented languages in the world!

"According to the World Atlas of Languages' methodology, there are around 8324 languages, spoken or signed, documented by governments, public institutions and academic communities. Out of 8324, around 7000 languages are still in use." https://en.wal.unesco.org/discover/languages

Most are at risk of extinction.

"half of the languages spoken today have fewer than 10,000 speakers and that a quarter have fewer than 1,000 speakers" (https://en.wikipedia.org/wiki/Language_death).

"Today, on average, we lose one language in the world every six weeks. There are approximately 6800 languages. But four percent of the population speaks 96 percent of the languages, and 96 percent of the population speaks four percent of the languages. These four percent are spoken by large language groups and are therefore not at risk. But 96 percent of the languages we know are more or less at risk. You have to treat them like extinct species." (https://en.wikipedia.org/wiki/Language_preservation).

"Over the past century alone, around 400 languages – about one every three months – have gone extinct, and most linguists estimate that 50% of the world’s remaining 6,500 languages will be gone by the end of this century (some put that figure as high as 90%, however). Today, the top ten languages in the world claim around half of the world’s population." (https://www.bbc.com/future/article/20140606-why-we-must-save...).

Re: Meta AI announces Massive Multilingual Speech code, models for 1000+ languages

#90

Post by syntaxing »

qwertox wrote: post #68

I would like to use stuff like this as a side-project. Buy a Nvidia Geforce GPU and stick it into my 24/7 server and play around with it in my free time, to see what can be done. The issue with all these AI models is that there's no information on which GPU is enough for which task. I'm absolutely clueless if a single RTX 4000 SFF with its 20GB VRAM and only 70W of max power usage will be a waste of money, or really…

I would recommend a 3090. It can handle everything a 4000 series can albeit slightly slower, has enough VRAM to handle most things for fun, and can be bought for around $700.
Post reply on HN