cbcase

Thanks for the kind words, Brandon! Been a busy couple of months :)

Mostly this, though it's not so black-and-white. The paper discusses results from a DNN-HMM system (Maas et al., using Kaldi) trained on 2k hours, and it does provide a small generalization improvement over 300 hours.

Much of the excitement about deep learning -- which we see as well in DeepSpeech -- is that these models continue to improve as we provide more training data. It's not obvious a priori that results will keep getting better after thousands of hours of speech. We're exited to keep advancing that frontier.

For a single utterance, it's fast enough that we can produce results in real time. Of course, building a production system for millions of users might require just a bit more engineering work...

Hi Jerome, those are great results! We got an email this morning from someone else on the Watson team pointing out that we didn't include the latest IBM number -- we'll be sure to update the results in the next version of the paper (three cheers for arXiv).

Of course, we openly say in the paper that we don't have the best result on easy subset of Hub5'00 (we had it as 11.5%). We're more interested in advancing the state of the art on challenging, noisy, varied speech. Of course we'll be working to push the SWB number down too :)

As in many things, it's a combination of both. For example:

- We wanted no more than one recurrent layer, as it's a big bottleneck to parallelization.

- The recurrent layer should go "higher" in the network, as it's more effective at propagating long-range context when using the network's learned feature representation than using raw input values.

Other decisions are guided by a combination of trial+error and intuition. We started on much smaller datasets which can give you a feel for the bias/variance tradeoff as a function of the number of layers, the layer sizes, and other hyperparameters.

I should add that I had the opportunity to work on this project and am happy to answer questions.

Glad to hear it! I had a lot of fun coming back to campus. Good luck to you too!

carl @ sift

The first two classes (106A and 106B) are very techinical, though I'd hesitate to let anyone who hadn't taken classes beyond them to work on a piece of software I had control over.

I think a good metric for classes is the final assignment, since it captures "how far" the class goes. For reference, the final assignments are:

106A --- It varies, but has recently been a text-based "Adventure"-style game (in Java) that requires tracking the map, player state, various objects and their capabilities, etc. I think there might also be a small graphical component.

106B/X --- Again, it varies, but the best assignment (in my view) is a BASIC interpreter that implements both a REPL and stored programs. All C++, it's about building a big list of abstract expressions of different types (assignment, etc) that can be executed by walking that list (taking advantage of dynamic dispatch) and tracking global program state. It's a nice intersection of data structures and (very simple) recursive descent parsing.

107 --- A heap allocator to implement malloc(), realloc(), and free(), written on top of mmap(). A fantastic assignment.

So yeah, they're not trivial classes.

It's not exactly paradise. Because classes are only ten weeks total (slightly less in spring), you can't exactly spend the first week or two shopping around without doing non-trivial work for every single class you're shopping --- an approach which doesn't scale, to say the least. There's still a sprint the first few days. But still, it's nice.