Term Weekly Goals

February 25

So the plan is to build a very simple numerical NN in Python that can hopefully run on my computer without a large data set. I want to create a basic problem and data set with known solutions and build an NN that can also predict those solutions -- the most basic of the basic.

I also want to look into pre-processing speech data and what I have to look forward to in the future regarding that. But I don't want to use speech data quite yet. I also want to see if I can get Tensorflow working on a lab computer. As it currently stands, it's incompatible with my home computer and I need more access to actually install Tensorflow so I'll have to send out some emails.

By keeping it bare bones, I can see just how long it takes for the NN training to take place. I can test how many iterations of testing I should eventually do for accuracy. And maybe I can try using a simple outside data set. That's the stretch goal though.

March 4

I want to begin using pre-processed speech (Tensorflow/PocketSphinx/SpeechRecognition) to see how long it takes for training to work on those models.

If I can, I want just the basic, "can this NN recognize the numbers 0 to 9 spoken outloud. I believe there's a dataset on TensorFlow for this but I'll have to see if I can get Tensorflow to work to begin with on this end.

If not, test around with PocketSphinx or SpeechRecognition.

March 11

This time, try to use live microphone data with an NN to see if it can read the numbers zero to nine again. This should use either the tested network from the previous week or maybe even a new data set. If I can start building up towards word recognition than do so. Hopefully, I'll have permissions to use the grad lab and maybe get Tensorflow working at this point.

If things start going south, it might be more beneficial to looking into how NNs currently run inside game engines to help smooth out movement. It maybe not be audio data, but seeing how they run in engine can also be beneficial to the overall project. Or it might be helpful to look into MATLAB as another method of programming an NN and weighing the pros and cons of Python vs MATLAB.

March 18

Try to use SpeechRecognition to understand speech. The goal is to have an NN that can break down phonemes. I think this will be the main goal of this term. If I can get phonemes recognized accurately that may be half the NN problem solved. However, the goal isn't to get words accurately understood. It's to get the mouth shapes done correctly for future animation through neural networks.