Speaker Recognition using MFCC and GMM

I've run the system using the following for training: Speech data(NTIMIT) --> MFCC (feature extraction) --> GMM (modeling)

for testing:

Speech data(NTIMIT)--> MFCC (feature extraction) --> EM (scores)

the accuracy I am getting is 44% for 461 speakers. it was confirmed by 2 at least(1. Reynolds. 2. Patra) that running such system should give an accuracy of 60.8% for 630 speakers i have done lots of changes in terms of sampling frequency (mainly 8000 or 16000), number of MFCC cepstums, number of MFCC mixtures and iterations and the window size and that was the best percentage I could get.

I am using an MFCC and GMM codes which gave good result with TIMIT

advice would be really appreciated

[hanieh rafiee]

hi how use mfcc with gmm in matlab? please help me

Answer 1

[mamdouh]

i can tell that you should give more attention to the training data and the prepossessing step .... you can use PCA algorithm for dimensionality reduction and the class separability measure i think it may help

and if you can help me with code of the mfcc and the gmm i'll be thankful Regards.

[Shaikha Hajri]

PCA didnt work as expected and dimentionality reduction helps mostly in execution time and degraded the accuracy!

you can get the codes from the VOICEBOX: http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/voicebox.html

[VINAY]

Hi, You can use DTW algorithm for matching the speech files and speaker recognition.

Answer 2

[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