As others here have pointed out, the value now for GOOG is in building the best training data-set in the business, as opposed to just racing to find the best algorithm.
First, you take the huge input (because something like sound has a huge amount of data in it, similarly with images there are a lot of pixels) and learn a simpler representation of it.
The second problem of mapping these nice dense features to actual things can be solved in different ways, even simple classifiers can perform well.
This doesn't actually need any labelled data. I just want to learn a smaller representation. For example, if we managed to learn a mapping from bits of audio to the phonetic alphabet then our speech recognition problem becomes one of just learning the mapping from the phonetic alphabet to words which is a far nicer problem to have.
Some ways of "deep learning" solve this first problem (of learning neater representations) through a step by step process of what I like to refer to as laziness.
Instead of trying to learn a really, really high level representation of your input data just learn a slightly smaller one. That's one layer. Then once we've got that we try and learn a smaller/denser representation on top of that. Then again, and again, and again.
How can you learn a smaller representation? Well a good way is to try and get a single layer to be able to regenerate its input. "Push" the input up, get the activations in the next layer, run the whole thing backwards and see how different your input is. You can then use this information to tweak the weights to make it slightly better the next time. Do this for millions and millions of inputs and it gets pretty good. This technique has been known about for a long time, but one of the triggers for the current big explosion of use was Hinton working out that this back and forth only really needs to be done once rather than 100 times (which was thought to be required beforehand).
Hinton says it made things 100,000 times faster because it was 1% of the computation required and it took him 17 years to realise it in which time computers got 1000 times faster. Along with this, GPUs got really really fast and easier to program. I took the original Hinton work that took weeks to train and had it running in hours back in 2008 on a cheap GPU. So before ~2006 this technique would have taken years of computer time, now it's down to minutes. Of course, that's then resulted in people building significantly larger networks that take much longer to train but would have been infeasible to run before.
But you still need a lot of unlabelled data. While I doubt google is doing that with this setup, they have done something before, where they setup a question answering service in the US that people could call I think for free to collect voice data.
TL;DR
You need labelled data. But it turns out you can learn most of what you need with unlabelled data, leaving you with a much simpler problem to solve. That's great because labelled data is massively more expensive than unlabelled data.
But, assuming that was their plan, they'd have a couple options:
- Like you said, they could turn it into supervised training examples by transcribing it. I'm sure they'd at least like to transcribe some of it so that they can measure their performance. Also, while Google does have a lot of 1st party applications feeding them training data, customer data might help them fill in some gaps.
- They might also be able get some value out of it without transcribing it. Neural networks can sometimes be pre-trained in an unsupervised manner. One example would be pre-training the network as an autoencoder, which just means training it to reproduce its input as its output. This can reduce convergence time.
Or, at least, that's my best guess with zero research and little knowledge.
Like if neural networks trained with user data should be un-copyrightable, and public domain by default.