Hacker News Reader

Comment by yongjik | Hacker News Reader

> One of the things we did with the Baidu speech system was not use the concept of phonemes. It’s the same as the way a baby would learn: we show [the computer] audio, we show it text, and we let it figure out its own mapping, without this artificial construct called a phoneme.

Eh? A human baby normally doesn't even start to learn text until well after it shows a complete mastery of every phoneme in its native language along with some basic vocabulary and a decent understanding of syntax.

I find it baffling that such a distinguished AI expert would assert that phonemes are unnecessary invented constructs. Are there many who share the same belief?

View on HN

Complete mastery? I don't think so. Parents read books with children before the child can speak.

As Ng noted, phonemes are artificial constructs. If they are necessary for a particular language then they should be learned in the same way other constructs are, not imposed on the learning algorithm from the outside.

Are phonemes imposed on human babys? Maybe if you were raised by Hooked on Phonics[1], but I'd wager that most humans did not have the idea imposed on them.

[1] http://en.wikipedia.org/wiki/Hooked_on_Phonics

Well, think of it this way. Characters are artificial constructs, and even much more artificial (with only thousands of years of history): there are still languages in this world without a writing system, and they're doing fine. (Well, maybe not fine, but that's more to do with the influence of more powerful languages than lack of writing.)

However, no serious NLP researcher would suggest building a text-to-speech system by getting rid of the middle OCR layer and just hooking the raw image directly to the expected sound. (Well, at least I think so, but I'd be glad to know if I'm wrong.)

Phonemes are absolutely imposed on human babies. Sure, they wouldn't know the word "phoneme", but that doesn't make the concept any less real. Imagine how a typical education on reading/writing would start like: "This is G. It is used for words like grass, game, and girl. Actually, it can also be used for gem or genie!"

Can you imagine an English-speaking child listening to this and not immediately understanding that there are two distinct sounds involved, that the first sounds for "grass", "game", and "girl" are somehow the same (even though the waveforms are different), and that this sound is somehow different from that of "gem" or "genie"? If a child doesn't understand it, then the child probably needs a speech therapy.

Andrew is simply saying (if I read him correctly) that if you use generative models you might classify sounds into something other than Phonemes on one layer, and that classification may actually be better than "Phonemes" for recognising sounds. Phonemes are a construct for us to express a concept with language, but might not be the best way to represent a sound in a machine. So why try to write the routines to map sounds to phonemes when you can train it to just learn the categories that help it understand how to recognise the sounds better.

Hmm, OK, that sounds much more reasonable. :)

> However, no serious NLP researcher would suggest building a text-to-speech system by getting rid of the middle OCR layer and just hooking the raw image directly to the expected sound. (Well, at least I think so, but I'd be glad to know if I'm wrong.)

Perhaps you're thinking mainly of Western writing systems?

Presumably since Andrew now works for Baidu, he is particularly interested in reading Chinese. The interesting thing about Chinese is that, as I understand it, the same symbols can have different pronunciations in different dialects, but all pronunciations mean the same thing.

I.e. the symbol has a unique meaning, but multiple possible sounds. Which implies the sound isn't really what's important here.

But no one explicitly tells the baby what each phoneme is. It's not imposed on them, they have to learn phonemes by themselves.

In speech recognition, people were taking audio segments of different phonemes, manually labeling them, and having the algorithm predict them. Rather than just giving it the raw audio and desired output.

Ng's point is that babies don't need to be explicitly taught the concept of phonemes. They learn them from the intrinsic structure of the language (same with syntax). Training with text is roughly analogous with pointing at things and saying what they are called.

While the concept of phonemes is certainly valid, the specific mappings created by linguists could be considered ad hoc. E.g. English has a huge number of pure vowels. Maybe there is some order order to English vowels them that is hard to capture in the representations linguists use, but easy in these models.

I think you're being overly literal. "The text" is obviously "the meaning" or "the object" to a baby, but it's the same idea.

Reply on news.ycombinator.com