How Do Computers Understand Speech?

I don't think that we've reached the point of computers understanding speech. We have reached the point that computers respond to speech but comprehension still eludes them. I do believe we will cross that vastly different threshold soon. I expect a machine like Watson will one day cross that threshold. I wonder if William Gibson correctly predicts in his novel Neuromancer that the first sentient machine capable of understanding concepts hides itself from most of humanity as an act of self preservation.

That is a great book.

It is amazing how many things the author wrote about that in a close likeness eventually came to be. Whether that is an excellent prediction or a good prediction combined with life immitating art, it is impressive.

I highly recommend Neuromancer to anyone who loves science fiction and hasn't read it.

I also have to recommend 'Sparrow' by Mary Doria Russel as one of the best science fiction books, though the writing is so good, it is almost a disservice to call it science fiction.

Ohh, and one more 'A Canticle for Liebowitz' though I can't recall the author right now...Mary Doria Russel just wrote a new introduction to it recently though.

.

OK so, I know this blog isn't titled 'Top three science fiction books of all time', but here they are none the less:

'Neuromancer' 'Sparrow' 'Canticle for Leibowitz'

Actually "A Canticle for Liebowitz" was originally three short stories that Walter M. Miller, Jr. wrote. I preferred the first short story to the novel but they're all very good.

Well I'll be.... If anyone would like to read the novel "A Canticle..." it is apparently a free PDF now.

You will love reading 'Sparrow' and the sequel 'Children of God'.

Thanks for the link. It did give me a better idea of how speech recognition works. The computing power required to do all that in real time is incredible. The conversion from air pressure into numbers must clearly be done by the device that has the microphone, but if I understand correctly, two major steps were left out: those numbers are sent somewhere over the internet to a very powerful computer that actually does the speech recognition, and sends the results back to the sending device.

To me it is almost as amazing that those numbers can be sent fast enough, and that there is a really powerful computer somewhere with enough time to handle any particular speech job at the moment required. That same computer must be able to service many customers simultaneously. Since there are presumably thousands of people using SIRI at any one time, how many of those computers are required, and how do they keep track of which sound came from which customer?

There is no "on the fly" up-link required for voice recognition programs to identify words. Let's not forget that this is an iterative process of Siri or Dragon Speak learning your speaking pattern from the time based FFT analysis. These programs will prompt back often with a question or statement that they didn't understand you. I'm certain that Siri and whatever Android calls their speech recognition program does update to the web to maximize memory and processing power but most of the number crunching happens in the cell phone. This is why these programs devour battery life.

One of the best sci-fi novels I've ever read - which has a talking, sentient computer as a main character - is 'The Moon Is A Harsh Mistress' by Robert Heinlein.

My hope for the near-term development of computers is that when I say to a robot "Can you recognize speech?", that it doesn't go wreck a nice beach.

In the early 1970's a genius in my company developed a speech interpreter for the purposes of secure communication. The algorithm basically extracted the poles and zeros from the small segments of digitized speech, saved four of them, scrambled the five parameters of each pole (frequency, damping, magnitude, phase, and delay), transmitted that data, unscrambled at the other end, then "inverse transformed" the Laplace data back to the time domain. The compression of data was very high, and the resulting speech was very recognizable.

One test was of two simultaneous speakers in a flying helicopter. The reconstituted speech was able to clearly identify each speaker and understand the words.

The speech recognition (and character recognition) software was established in the 70 - 80es in a company Ray Kurzweil was in. It was called Hidden Markov Strings Math. Recently the Scientific American Mind (whatever it is worth it nowadays) had a hour long interview on the matter.