Baidu Releases Open Source Artificial Intelligence Code

An anonymous reader writes: Chinese web services company Baidu has released a new artificial intelligence software called WARP-CTC. The code is apparently capable of speech recognition, particularly for short segments, that exceeds human capability. The source code uses an approach called 'connectionist temporal classification' and has been released on GitHub.

"better than human" was achieved in 1994, already

[ffkom]

I've been developing automatic speech recognition systems in the 1990s. Back then, the best performing recognizers were based on Hidden Markov Models, and for "out of context" tasks like "determine whether an individual spoken word from an unknown speaker is 'nine' or 'none'", the automatic recognizers already achieved better recognition rates than humans. However, the specific human strength when recognizing fluent speech is to (a) quickly adapt to different speakers and (b) to fill in all the uncertain words from the understanding of the context, requiring "world knowledge". And that strength makes a very big difference. So the claim in the article is not really anything special, it is to be expected that computers are better than humans in this special task, for at least the last 20 years.

Re:"better than human" was achieved in 1994, alrea

[Gravis Zero]

I've been developing automatic speech recognition systems in the 1990s. Back then, the best performing recognizers were based on Hidden Markov Models, and for "out of context" tasks like "determine whether an individual spoken word from an unknown speaker is 'nine' or 'none'", the automatic recognizers already achieved better recognition rates than humans.

Sure... but this time they did it with Mandarin instead of English.

Re:

[Anonymous Coward]

For those who may be interested learning more about a certain subject, here's an unlinkified URL of a video that may or may not contain any interesting information. I'm not going to tell you anything about the contents, let alone an interesting summary.

Re:

[Anne Thwacks]

Most of the current crap fails miserably with English - it may or may not work with American - I could not possibly comment.

Re:

[Rob Lister]

I had no idea they have been that good for that long. In the wild however, they suck. I hope I'm not coming off as super-critical. But it is the fault of humans for trying to make the machine pretend to be human.

I want to start this by saying that the following is decidedly NOT satire.

It is most probably the manner in which they are used that sucks the most. But maybe it is my own personal bias that is really to blame. Which is to say: I am prejudice against AI programmers.

When a robot answers

Re: "better than human" was achieved in 1994, alre

[Zalbik]

Take your pointless rambling nonsense somewhere else please, this is Slashdot.

Dang, this is Slashdot?!?! Never mind....carry on.

Re:

[nycsubway]

References please? I'm interested in this.

Good

[penguinoid]

Now, can they please put this code to use on those terrible phone menus? With single words or phrases, like the article says? They can even have the advantage of AI-understandable context -- only a few responses match the question -- and yet they still get it horribly, hilariously wrong.

Re:

[rtb61]

Artificial intelligence do not belong in portable devices not for any paranoid reason but simply energy efficiency. To be actual artificially intelligent it must be capable of learning and investigating new solutions. So right or wrong cease for all non core functions where better or worse answers dominate. So logical, when queried the artificial intelligence will produce a 'good enough' answer, which is either used or rejected whilst in search for a better answer. Even with the good enough answer is used,

Re:

[jandersen]

The irony that this code is on GitHub is rich.

Not the only irony around here. When I grew up, everybody KNEW, they just KNEW, that those evil, Chinese Communists should open up their society, embrace Capitalism and take part in the international society, where we freedom-loving Westerners would welcome them and teach them how to think. So they did, and now they are being pummeled for that very thing, not least because they dare to make up teir own thinking. It's open source - you are supposed to be able to inspect the code, and in fact, one of the fund

communist vs capitalist

[Anonymous Coward]

Apparently this CTC technique was pioneered by the Swiss (IDSIA) who attempted to capitalize on it (Lifeware). Where the first impulse of the chinese researchers at Baidu was to post it on github.

To be fair, the Baidu researchers are located in Silicon Valley. So maybe this is just comparing socialists to capitalists.

Recognising certain phrases...

[Catmeat]

Is it smart enough to automatically call the secret police if it hears the words "Falun Gong".

What...

[liqu1d]

How on earth can you exceed human capability for understanding speech? Foreign tongue ok fair enough otherwise huh.

You would be surprised how bad you are...

[ffkom]

... understanding speech if you had somebody (not yourself, of course) make recordings of strangers on the street uttering just a single word, each, randomly picked from a dictionary book dictionary. Chances are your recognition rate would be somewhere around 90%.

Automatic recognizers achieve better rates on this task - but they'll loose against you when it's complete, sensible sentences that are being spoken, even more so if you heard more sentences from the speaker, before.

Re: You would be surprised how bad you are...

[liqu1d]

Ahhh ok. Clearly I didn't understand the process there. Thanks :).

Re:

[Anonymous Coward]

The Baidu code has 110 percent success rate --- the benefits of totalitarian capitalism revealed.