ID = voxsrc20_ask_08
Status: closed
Question
What is speaker space?
Is it 2D or 3D, and what is the axis?
Answer
Speaker space is high dimensional space, for example 256 or 512 dimensions.
It is highly abstract space, not just 2D or 3D.
In some visualization tool, 2D figure of small amounts of speakers can be selected
for better understanding of cluster distribution.
Keep in mind the following things:
- One speaker is one class
- One utterance of certain speaker is one object of this class
- The target of training is to get highly discriminative embedding that can have small intra-class distance
while large inter-class distance. - All utterances of certain speaker forms a cluster in speaker space.
- The size of speaker space is depended on the speaker number. The larger the number, the larger the speaker space.
Reference
[1] SpR_1706.xx000_A review of Speaker recognition, doi: 10.21817/ijet/2017/v9i3/170903513
[2] SpR_1711.00366_Full-info training for deep speaker feature learning