Speaker recognition

 I am going to do a project based on speaker recognition (not speech recognition). I saw many files in the internet and came across many methods.
First part of the program is that they find MFCC and after that we have to do the pattern recognition.

Here are some of the pattern recognition algorithms that I came across 1)VQ algorithm followed by LBG algorithm for clustering. 2)K means algorithm.

I have another idea. I know neural networks for pattern recognition in image processing. Will that work with speaker recognition?

Is there any code in matlab central for speaker recognition? If you have done this project before please tell me the method that you followed.

Thanks in advance.

Much awaiting for your response.

[Walter Roberson]

Note: this field is usually known as "voice recognition".

[Image Analyst]

From the subject line I thought he was talking about speaker identification (recognizing a particular speaker and extracting his speech), like the "cocktail party problem" http://research.ics.tkk.fi/ica/cocktail/cocktail_en.cgi but I've heard of that being solved with ICA, not the acronyms he listed.

Answer 1

[William]

Many use a Gausian Mixture Model (GMM) after using the MFCC. There is a really good toolbox for these operations called "voicebox.m" it is a collection of functions that all you to extract and classify data from speech via wavread()

[i Venky]

You mean the voicebox file uploaded in the internet?

Is GMM better than other methods? If you have done this speaker recognition before, please tell me how you did it.

Answer 2

[i Venky]

William I am much awaiting for your response.

Answer 3

[William]

Look over this website. I had to do this a year ago for a class and this is exactly what I followed

http://scgwww.epfl.ch/courses/Biometrics-Lectures-2005-2006-pdf/03-Biometrics-Exercise-3-2005/03-Biometrics-Exercise-3-2005.pdf

Here is the link for voicebox

http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/voicebox.html

[i Venky]

How was the accuracy?

[William]

It isn't terrible. if you collect a lot of data with the MFCC than your model might be more accurate but the time to process will slow way down.

Answer 4

[i Venky]

 Do I have to use the guassmix function of the voicebox?

Thanks in advance.

Answer 5

[William]

There are numerous GMM algorithms that could be used to do this. find one that you understand so that if it ever stops working you can figure out why.

[i Venky]

Thanks and I will notify you once I finish the program.

[kml]

did you finish the program?

Answer 6

[i Venky]

What do you think about neural networks? It worked perfectly when I was doing image processing. Will that work with this?

[Greg Heath]

Both the MLP and RBF with a single hidden layer are universal approximators and can be used for both regression and pattern

recognition. If you are familiar with the NN Toolbox, it

shouldn't take long to try both.

Greg

[i Venky]

Hello Greg Heath. I am not that much familiar with nn tool box. I used GUI for image processing. I usually use the neural network pattern recognition with a two-layer feed forward network window. I tried the same for voice recognition and it didn't work. So I think I should go for some other nn method.

In the nntoobox how would you select MLP and RBF?

Thanks in advance.

Answer 7

[William]

I don't have a lot of experience with Neural Networks. If you understand them try it and see what happens. Worst case you lose sa few hours of time.

[i Venky]

Hello William. I found out the mean of the log likelihood for the given input. After this how would you find the best match. There are about 12 means (12 components in the GMM) for every model. How would you compare the models based on the mean of log likelihood in this case. (If there was only one mean for every model then I would find out the maximum value of the log likelihood and choose the best match but here I have 12 means so I got confused.)

Thanks in advance.

Answer 8

[Brian Hemmat]

Audio Toolbox provides several examples for speaker recognition (both identification and verification):

• Speaker Identification Using Pitch and MFCC
• Speaker Verification Using Gaussian Mixture Model
• Speaker Verification Using i-Vectors