stephensonsco

Deepgram just posted an open sourced Kur Speech Recognition example based on Deepspeech (https://arxiv.org/abs/1412.5567) on our blog. We'd love to help people get it running and spread the DNN speech love.

For sure. Install Kur and follow the examples. There's no voodoo, just pure clean DL fun.

Great point about virtualenvwrapper. We've included that too in the installation instruction.

The install instructions are pretty much structured as: quick,

  pip install kur

cleaner and quick,

  install with vituralenv in a quick script

best practice but more setup,

  install a whole environment and include virtualenvwrapper

We use it and like it because we are trying a lot of models per unit time. And we want to change them, slide in new data, transfer weights, and others things like that, without headaches.

So the real reason is for internal efficiency for model prototyping, or more direct: results per human-hour. If we have to spend time in troubleshooting land, then we're losing our Startup Competitive Advantage™ in DL (the ability to move fast).

You'll be able to easily do that once we release Kurhub.com.

Oh yeah, Kur allows MUCH faster prototyping. And if you utilize Jinja2 (Kur supports that in the Kurfiles), you really start unlocking the time savings.

Thanks for bringing up other frameworks in DL! Amazon's DSSTNE is restrictive, but it's great for their purposes. On the DSSTNE GitHub you can see statements like:

"DSSTNE currently only supports Fully Connected layers ..."

Kur supports the cutting edge: like CNN/RNN.

"DSSTNE Engine works with data only in NetCDF format."

Kur supports the data that you have on hand. You can see in the tutorial (http://kur.deepgram.com/tutorial.html) how easy it is to send brand new data in the familiar Python pickle object.

These are the kind of GOTCHAs that people doing deep learning run into all the time. And they are a major time suck. Kur relieves you of those duties so you can work on more interesting parts, like trying novel models :).

We're so glad DL tools are coming out. But there's still tons of progress to be made and Kur is one step along that path—making the user experience more efficient and enjoyable.

Thanks! We really push to keep the dependencies clean so that all you need is Python 3.

Doing:

pip install kur

Installs all of the packages that Kur relies on, so you don't need a list to run down and install.

Anyway we can help get it installed for you?

About Windows: great point and we have had users test in Windows. It installs fully there too!

Try out the examples, tweak the models, and let us know how you get on!

Also, we're hanging around on gitter if you want real time advice :)

Deepgram is open sourcing Kur (http://kur.deepgram.com)! Kur is the worlds first descriptive Deep Learning software. Think of a model, describe it in a simple YAML or JSON file, and train to get state-of-the-art results. There's no need to code.

Why we built Kur: Prototyping DNNs is a slow process. Most people doing deep learning want to iterate and try out different model architectures and learn from others. It's hard to do this using barebones backends like TensorFlow/Theano or even the higher abstraction of software like Keras.

Kur is not speech specific. It can be used for images (we supply two examples), speech (we supply one example and DG is open sourcing a new audio dataset with it, the DEEPGRAM10), text, etc.

There are CNN layers, RNN, dense, dropout, batch norm, etc. to pick and choose from. The best part? Kur does all the plumbing! You want one input but two outputs? Not a problem, describe that model in Kur!

We're really pumped to be releasing Kur and would love to answer questions if you've got em.

Thanks! Deepgram AI Research Team

http://kur.deepgram.com http://github.com/deepgram/kur http://kurhub.com

Great point and we are definitely looking into this. How much is the right amount for a few seconds?

I'm thinking laid back party parrot.

debug mode is fully on!

We're working on the problem now. Everyone churned through audio faster than expected. New hams coming up.

Deepgram (W16) is pretty pumped about HamPanda and all the fun that SHOULD be had with it. Seriously, go have fun with it.

The magic is mostly brought out by drshrey, I'll have to get him in here.

Exactly!

I'm pretty pumped about this. Things are looking really good but still not obviously in the clear. Cautiously optimistic :)

Looks fixed now

Hey, I'm the blog post author (Deepgram cofounder too)! Sorry if the accuracy wasn't clear enough. The numbers quoted are for search accuracy, not for speech-to-text accuracy. It's a subtle thing but makes a lot of difference (and it is the thing that matters most when you are in trying-to-find-something mode).

If you ran phone calls or tape recorder audio through a speech-to-text engine then the word accuracy rate is like 10-50% (i.e. abysmal). When you try to search for keyphrases like "frolicking kitten" the likelihood of a text match with STT is ~20%.

If you ran that same search with Deepgram then 80% of the time you'd find what you are looking for since Deepgram doesn't have to guess at what is being said, it takes the inverse approach and matches 'how it sounds' using deep learning voodoo magic™.

I can't really add much more in the way of math. Linear algebra and calculus get you really really far.

It's true I didn't mention this but I was hoping the coverage in the related links would be sufficient. I would loved to have had a section that points out some basic tenets though.

You are right that most resources aren't that great at showing you the way. There is a lot of grinding to get past obstacles.

We are trying to address this soon!

Good point. Should probably say "you'll get an interview". I think you snipped out something pretty important though. The fact that you have done something nontrivial with DL is actually a big tell for the employer. If you just 'want to learn more about DL because it seems interesting' then employers aren't interested. But if you say "I am a good engineer and I have already done something nontrivial" then you have a much better chance of getting an interview and then the job.

If you:

- have good problem solving and coding skills - spend a month or two learning how to build networks using good libraries

Then you will be able to get a good result with the Twitter task I pointed out. It takes being able to work input data correctly, think about what matters and what doesn't, then synthesize using readymade DL tools, usually in python. None of that is complicated esoteric neural net stuff, it is just motivated problem solving.

I'll add that as a big point — you should be able to code, problem solve, and be motivated.

For the people who want a job in deep learning, they are definitely around and they will be for a while.

It's true that DL will probably become just-another-library, but that will happen only once computing becomes extremely cheap on the petaflop scale (it isn't cheap yet). Even after that happens, the people that spend time doing DL now will be trained in a way of thinking that will be in demand for a long time.

If you've got a lock on calculus, programming, and linear algebra, then you've got the skills to understand deep learning. Most time spent working on DL is not in the megamath part, it is in finding good network structures to optimize results.

These are definitely great points. Most companies looking for DL/ML talent aren't interested in setting up HR hoops for the applicant to jump through.

They want to see if you did cool stuff before you applied for the job. If you didn't then you won't get an interview, but if you did then you have a chance no matter what your background is. Of course, the question of "what is cool stuff?" comes up. If it is building small projects with a a little bit of success, that probably won't do it (it might work for larger companies, or companies that need light ML/DL performed). But if it is "built twitter analysis DNN from scratch using Theano and that can predict the number of retweets a tweet will get: here's te accuracy, here's a link to a my write up on it and here's a link to github for the code.".

Edit: added words similar to this at the end of the blog post.

If you are a great software engineer that can solve hard problems yourself, and then you add to that real experience solving your own nontrivial problems with DL, then people will hire with no questions asked. The reason is that not many people have experience in DL. So companies need to get good engineers that have the problem solving mindset and motivation+creativity to learn new frameworks and lick the problems they face in new awesome ways. Industry is dying for people like that.

If you are that type of person then you can kill it with just a few weeks of hard study on your own.

You're definitely right that most ML jobs aren't DL all day everyday. But if you are working with data that is rich and fairly homogeneous but hard to model (like recorded speech) then you'll probably benefit a ton if you use DL due to the ability to learn underlying representations. If that's what you are up to then you do spend all of your time building and tuning DNNs (all three ML people at Deepgram are doing that at this very second :) )

good catch. fixed it

just wrote a blog post that I think a lot of folks will like if they are looking for a job in ML/DL.

Would love to hear if I missed something!