1. WHY --- 为什么需要使用GMM-UBM来建立Individual Speaker Modeling?
"Usually, we do not have much data from a single speaker. In most practical cases related to text-independent scenarios, the enrollment data is at best in the order of a minute." “Assuming that the number of Gaussians used in the mixture mode, Γ ∼100,
then it is easy to understand why there is no where close to enough data to be able to estimate the mixture parameters for the speaker.”
2. WHAT --- UBM模型和Individual Speaker Modeling长什么样?
- the GMM are usually diagonal.
- Practice has shown that it is advantageous to train two separate background models, one for female and the other one for male speakers.
- the number of Gaussians used in a UBM是多少?
3. HOW --- 怎么基于UBM模型来建立Individual Speaker Modeling?
- 调整什么参数?
Only means adapted 或者 All parameters adapted
- 用什么方法调整?
通常是MAP方法,“For very short enrollment utterances (a few seconds), some other methods have shown to be more effective. Maximum likelihood linear regression (MLLR) ”
4. Advanced Topic
- 怎么减少GMM-UBM的computationally intensive
- 怎么解决GMM-UBM的phonetic mismatch problem
参考:
[1] An Overview of Text-Independent Speaker Recognition: from Features to Supervectors