In fact, I didn't work on these methods but I know that i-vector is an enhancement of the GMM-UBM method.

GMM-UBM: M=m+Dz

i-vector: M=m+Vy+Ux+Dz

Please, read this compact explanation:

http://cslt.riit.tsinghua.edu.cn/mediawiki/images/c/cb/131104-ivector-microsoft-wj.pdf

or this detailed one:

https://www.cs.swarthmore.edu/~turnbull/cs97/f09/paper/reynolds00.pdf

Thanks,

Actually I am already working on i-vector extraction, the equations are your wrote are some how different.....

i-vector : M=m+Dz and 

joint factor analysis M = m + Vy + Ux +Dz

with mean supervector.....

Thanks Hamid Eghbal, Yes I am already using this toolbox, and some how understand theory behind this, but still, most the the technologies they are using GMM-UBM still....

Dear Rizwan, we have to keep in mind the I-Vector and novel machine learning paradigms are associated with the nature and availability  of databases,  cannonical i vector may give the same or worser performance compared to GMM-UB under small databases without session variability or text constrained context or even short duration utterences , all these issues are active research areas.

@Messaoud Bengherabi. I have tested both GMM-UBM, and i-vectors, and in my experimentation, GMM-UBM gives better results in comparison to i-vector, specially on short utterances. 

Dear Rizwan

it may be a valid conclusions, but you can tune some parameters in ivectors or use duration as an extra uncertaintity  in building your i vector models or in scoring stage and I vector systems will be competitive or better than GMM-UBM.

@Messaoud: Yes, i-vector should produce better results, but I tried different combinations (short, long duration), and still the GMM-UBM seems produces better results. i.e I am testing my system with both (phone and microphone data). If I have only phone data for user, the system works only for phone data, and vice versa (microphone). But if I have both phone and microphone utterance for enrollment, then again GMM-UBM and i-vectors works, but this issue is that if I have only phone utterance, and test with microphone, (i-vectors should work) but it is not working. 

Hello all,

Can someone help me with what are the advantages and disadvantages of GMM-UBM and i-vectors over each other?

Every help will be appreciated.

Thank you.

Bhushan Muthiyan I think the paper given below might give you some insights about your question.

https://ieeexplore.ieee.org/document/8081264