Question about speech conditioning

[asr-pub]

In the paper :

Two encodings were produced for each training sample, which are averaged together.

Why use two encodings (one is self, the other is another clip of the same person speaking) ? How about use the single encoding of current sample ?

[manmay-nakhashi]

because if we just use the same audio as conditioning it overfits into that data.

[asr-pub]

So, during each training iteration, for the same audio, its Speaker Condition varies, which means that SC is dynamically computed ?