Audio-AGI / AudioSep

Official implementation of "Separate Anything You Describe"
https://audio-agi.github.io/Separate-Anything-You-Describe/
MIT License
1.64k stars 118 forks source link

how to construct training sample pairs? #9

Open anonymoussky opened 1 year ago

anonymoussky commented 1 year ago

Nice work!

How do you construct the training pairs? It looks like you construct the training pairs in the SegmentMixer class. Do you use the same minibatch sources to construct the "mixture" and "target audio" pairs?

There might be one issue: source1: male speech , s1 source2: another male speech, s2 if mixture = s1 + s2: as both captions are "male speech", will it confuse the model training?