Simple neural network question

[GSR]

Hi,

I am reading a neural network book and try to learn it by myself. I have some idea how neural network could use to classify time-independent patterns (like classify between 2 letters). However, I have no clue how it can classify a time signal. For example, I want to classify whether a song is singed by a male/female singer. How to do that?? I need only general conceptual suggestion. Any thing like websites suggestion will help

Thanks

Edited June 24, 2010 by GSR

[Francois Normandin]

There used to be a NI Labs: Neural Nets API, but I can't find it anymore.

NI Labs page doesn't mention it and all I can find on Google is an 18-months old discussion forum.

I still have the ZIP file on my PC.

PM me if you'd like to have it. It's > 5MB, so I'd rather not post it here.

[EricLarsen]

Hi,

I am reading a neural network book and try to learn it by myself. I have some idea how neural network could use to classify time-independent patterns (like classify between 2 letters). However, I have no clue how it can classify a time signal. For example, I want to classify whether a song is singed by a male/female singer. How to do that?? I need only general conceptual suggestion. Any thing like websites suggestion will help

Thanks

You've really hit on the $64 million question in neural net programming. Pre-rocessing the data into a discrete form that can be trained into a neural net is the most important factor in determining a net's success. There really isn't a one size fits all approach that will work for all time variant signals. In your specific case, the most likely difference between male and female voices is frequency. You might try doing an FFT on portions of the song and dividing that up into discrete frequency bins. The relative values of the bins are then fed to the input nodes of the net. Doing a search on neural networks and voice recognition will result in thousands of hits, one of them might be useful.

Here's a few things I've learned the hard way. Always normalize your input data set, variation is signal strength will really throw off your results. Make sure you randomly select values out of your training set for your test set. And collect as many training sets as you can.