The Gaussian Mixture Model with Universal Background Model (GMM-UBM) takes individual class training data to train the GMM and uses all training data or different dataset to build the UBM. In this way, it can make a fast and speaker or speech or class-dependent model for each individual. However, the GMM-UBM classification rate is lower than the Support Vector Machine (SVM) as I found in my studies. Noted, the SVM is a supervised classifier and kind of neural network. The SVM takes all training data at a time to make the training model and takes huge time to compute cost and gamma parameters. But, it gives a better classification accuracy than the GMM-UBM. Now, my question is why people use GMM-UBM for most cases?
In data analysis non of the classifier is best for every database. That is there is no general classifier so for which is superior for every database. Consequently, we cannot reject or ignore some classifiers accept the others in general. In fact some classifiers are good recipe for one database than other for the different ones. Also every classifier have their own pros and cons. For example SVM has a benefit to use all the training data but very time consuming, on the other hand Bayesian classifier is lightning fast but uses small samples for training and testing. Same is the case for GMM-UBM.
So, is this only data set issue? What about features? The GMM uses mixture component independently and it is expected the input feature will be decorrelated. Thus a feature with DCT gets benefit when GMM-UBM is used as a classifier.
Data normalization, conventionality reduction, filtering, militarization or data smoothing are usual data per-processing techniques. The issue is tractability between model and data. The model may be parametric, or the non-parametric or supported by SVD or Convolution etc,. or whatever, but the important thing is best coherence between method/algorithm/model/technique and given data at hand.
A good book for such analysis is attached.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification