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

High-performance speech recognition with no supervision at all

ai.facebook.com

Post reply on HN
Print view View original

41–50 of 66 posts

Re: High-performance speech recognition with no supervision at all

#41

Post by unityByFreedom »

strikelaserclaw wrote: post #40
unityByFreedom wrote: post #29

That's incredible. Since translation can be done this way too, it may truly be a short time until widespread use of babel-fish tech becomes reality. What a world!

Babelfish like technology would be a dream come true for me. At this point in time, google translate speech recognition works well when the speaker speaks a bit slow and uses simple sentences. The translation is the same, works well for simple sentences even though the translations in my language sound weird because it uses the equivalent of old English. There is still a way to go before speech recognition can recogn…

> There is still a way to go before speech recognition can recognize fast speech with colloquialisms, and probably even a longer way to go before it can translate longer sentences with abstractions into something that is clear and concise in the target language.

Did you watch the video tweeted by Facebook's CTO?

https://twitter.com/schrep/status/1395766932104572928

The speed, accent, and word choices of the speaker would throw off previous state-of-the-art tech.

The amazing thing about this new unsupervised algorithm is they can throw unlimited amounts of unlabeled audio and text at a computer and wait for it to become great. It's not AGI but it is still amazing. Language detection is also already solved so there really is no labeling required.

Re: High-performance speech recognition with no supervision at all

#42

Post by nharada »

renox wrote: post #34

If it's unsupervised why English transcription is much better than the other languages?

It looks like many of the components used are still English biased, for example the off the shelf tool they used to generate phonemes from text was a state of the art system for English and a more general but less performant version for other languages.

Re: High-performance speech recognition with no supervision at all

#43

Post by tjpnz »

Causality1 wrote: post #5

Will Facebook license this tech out? I can't be the only one who's noticed Google's speech recognition has gotten significantly less accurate the last couple of years, probably the result of some cost-cutting strategy.

There was a time a few years back where Google's speech recognition was virtually flawless (at least for my voice) but nowadays it verges on useless. I've had similar experiences with Google Maps (specifically navigation) and Google Translate. My understanding is that the implementation of these products is now mostly ML driven whereas in previous iterations ML was just one aspect of it. Curious to know how much truth there is to that or if I'm just viewing things with rose tinted glasses.

Re: High-performance speech recognition with no supervision at all

#44

Post by astrange »

joering2 wrote: post #30
joeyspn wrote: post #19

Earlier quoted context omitted.

Haven't you ever wondered if some mobile apps listen to you? And then you get accurate targeted ads...

People keep downvoting me, because "we need you to send a proof", but sorry I don't want to share Snowden's fate, but I know for fact IOS/Android, most TVs and major TV cable boxes (USA) are indeed analyzing your speech and sending back home _keywords_ from your conversations. I can imagine that's how they avoid being sued into oblivion for clear 4A violation, as metatags have been considered not a violation of your…

They ain't doing shit. Voice assistants only listen for their built in trigger words, if you even turn that on, and they do what they say with the conversation afterwards.

4A doesn't matter anyway, what matters is GDPR, it's not worth building a system that is not GDPR compliant even if you're outside the EU.

Re: High-performance speech recognition with no supervision at all

#45

Post by contriban »

sirfz wrote: post #4

In the title image, the Arabic word "ahlan" (welcome) is spelled correctly but left to right instead of right to left (should look like this: أهلا)

It takes effort to screw up a single word while copy-pasting it from a Google search result page.

Re: High-performance speech recognition with no supervision at all

#46

Post by contriban »

jsemrau wrote: post #16

Why would FB work on such a tech?

Are you looking at Facebook with glasses from 2010? Because one would expect a almost-trillion-dollar company to work on more than a single-focus product.

Why would a search engine (1997) make glasses (2013)?

Re: High-performance speech recognition with no supervision at all

#47

Post by KaiserPro »

joering2 wrote: post #30
joeyspn wrote: post #19

Earlier quoted context omitted.

Haven't you ever wondered if some mobile apps listen to you? And then you get accurate targeted ads...

People keep downvoting me, because "we need you to send a proof", but sorry I don't want to share Snowden's fate, but I know for fact IOS/Android, most TVs and major TV cable boxes (USA) are indeed analyzing your speech and sending back home _keywords_ from your conversations. I can imagine that's how they avoid being sued into oblivion for clear 4A violation, as metatags have been considered not a violation of your…

Ok, for those that want proof, its pretty simple to do.

1) we know that sending voice data to "HQ" costs power

2) we know that live transcription costs a huge wedge of power

3) we know that wakeword matching is quite power efficient. (see https://rhasspy.readthedocs.io/en/latest/wake-word/, https://github.com/MycroftAI/mycroft-precise)

So, in a quite room we know that to save power and data, devices won't be streaming data/listening. We can use that as a baseline for power and network usage.

Then we can start talking, measure that

then start saying watchwords and see what happens.

That aside, we know that its really expensive to listen & transcribe 24/7. Its far easier and cheaper to monitor your web activity and surmise your intentions from that. There are quite a few searches/website visits that strongly correlate to life events. Humans are not as unique and random as you may think, especially to a machine with perfect memory.

Re: High-performance speech recognition with no supervision at all

#50

Post by xiphias2 »

pesenti wrote: post #35
skybrian wrote: post #14

How does Facebook use speech recognition in its own products?

Our main usages of speech rec are: - Understanding for videos for content moderation, ranking and recommendations - Captioning - Assistant tech in Portal, Oculus and upcoming glasses

For me the most scary part is not doing end-to-end encryption in Messenger, which means that the US government has a right to listen to my conversations on it.
Post reply on HN