How Baidu is improving Mandarin voice recognition with deep learning
medium.com
medium.com
There is some data about this claim here: https://gigaom.com/2014/12/18/baidu-claims-deep-learning-bre...
Anyone here that has actually used their system and compared it to others?
The data looks impressive, but I am a bit skeptical of this claim, because I compared the Baidu OCR API for Chinese character recognition with the one from Microsoft recently. And I was disappointed. In many cases they were just as good as Microsoft's litte OCR dll that ships with Windows 10 [1]. Of course, OCR and speech recognition are somewhat different fields, but if anyone is an expert for Chinese language, then it should be Baidu.
It's not about being experts in Mandarin. The basis of their approach is that it doesn't encompass any expert design. It's an end-to-end deep learning approach. From the article:
> Our system is different than that system in that it’s more what we call end-to-end. Rather than having a lot of human-engineered components that have been developed over decades of speech research — by looking at the system and saying what features are important or which phonemes the model should predict — we just have some input data, which is an audio .WAV file on which we do very little pre-processing. And then we have a big, deep neural network that outputs directly to characters. We give it enough data that it’s able to learn what’s relevant from the input to correctly transcribe the output, with as little human intervention as possible.
> One thing that’s pleasantly surprising to us is that we had to do very little changing to it — other than scaling it and giving it the right data — to make this system we showed in December that worked really well on English work remarkably well in Chinese, as well.
So they've quickly trained their Deep Speech engine [1] to process Mandarin after first training it to transcribe English, without injecting specific language expertise into the engine.
Finally, I strongly doubt the OCR and speech recognition teams are the same. I don't know about the OCR team but their speech recognition team is based in California [2] and includes Andrew Ng and Awni Hannun from Stanford University.
[1] http://arxiv.org/abs/1412.5567
[2] http://usa.baidu.com/deep-speech-lessons-from-deep-learning/
This is fascinating. I've been living in Vietnam for a while now and struggling to learn Vietnamese. Mastering any tonal language is quite difficult for native English speakers so I've wondered how hard it would be to develop speech recognition algorithms for Chinese or Vietnamese. Sounds like it's actually not that difficult but it requires a different approach than non-tonal languages.
It doesn't seem to make any sense to have the same words mean different things if you can just make a different word...
Sarcasm aside though, a whole lot of languages are tonal. Most languages spoken in Asia and many African ones are. And you use tone too: what's the difference between "You there." and "You there?" and "You there!". It just serves a different purpose in English than it does in Chinese so it seems less mysterious to an English speaker.
The only tonal language I can really speak about is Mandarin (although I encourage you to read about other systems). The way the tones work in Mandarin isn't as complicated as it sounds. The classic example is this[1]:
1. mā (妈, mother). This sounds like a somewhat high-pitched, steady tone
2. má (麻, hemp). This is called a rising tone, and it sounds sort of like the end of a question in English
3. mǎ (马, horse). This tone falls and then rises (sort of like a #4 then a #2). Imagine a whiny child saying "mooooom!" to get their mom's attention
4. mà (骂, scold). This tone falls quickly, it sounds like in the movies when someone shouts "Hey!" at a robber that's getting away
5. ma (吗, an interrogative particle). This is a neutral tone, its sound depends mostly on the syllable before it but in isolation it either falls slightly or sounds like an English "toneless" syllable.
Mandarin only has those 5 tones, it's not like "ma" can sound fifteen different ways. They aren't differentiated by the pitch itself exactly, it's whether the pitch is rising or falling.
A Chinese speaker might describe the difference between má and mà the same way you'd describe the difference between "bet" and "bit" to a speaker of a language that doesn't differentiate those vowels. Sort of like rhyming, but not quite. As an easier example, Spanish doesn't really have a contrastive /ɪ/ as in "bit", so a Spanish speaker might have the same trouble differentiating or producing it without some practise that you might have doing the same to má or mǎ.
[1] https://en.wikipedia.org/wiki/Tone_(linguistics)#Mechanics
Is Chinese missing vowels?
Usually complexity increases because it confers some kind of evolutionary advantage.
Language doesn't follow the same rules as biological evolution. For instance, prestige dialects often cause people to opt to use a less useful language for social status reasons. Pidgins can create amalgamated languages with essentially random features of the parent languages. Many language changes really confer no "benefit" at all, like a change from agglutination to strict word order (and back again).
It's the other way around: http://wals.info/feature/13A#2/28.6/152.8
And there is some evidence that there have been shifts in tonal complexity within languages over time [3].
[1] https://www.mpg.de/8879447/tonal-languages-institutes
One you have "just" made your different words, then you can "just" teach them to one billion or so people, and you are done.
What does make sense is that languages evolved over long periods of time. People didn't "just" make different words that everyone suddenly used. Just as English has "digest your food" and "Reader's Digest" with different intonation, Chinese distinguishes meanings and words with different intonation as well. Not exactly the same because there's also a difference in the written form in most cases. It doesn't have to make sense... it just is.