Choral Music Separation Demo

Improving Choral Music Separation through
Expressive Synthesized Data from Sampled Instruments

In this demo page, we present two parts:

First, we present the synthesis demos from SoundFont-Synthesis, our proposed Standard-Vocal and Expressive-Vocal datasets.

Second, we present separation demos from real choral music datasets of three models:
(1) None-dataset pretrained models; (2) SoundFont-Synthesis pretrained models; and (3) Our Expressive-Vocal pretrained models.

Synthesized Choral Music Dataset Demos:

Demo

SoundFont-Synthesis

Standard-Vocal (Ours)

Expressive-Vocal-Vowel (Ours)

Expressive-Vocal-Word (Ours)

#1:

#2:

#3:

Short Separation Demos (from Bach and Barbershop Collection Dataset):

Demo #1

Mixture:

Voice

References

Non-dataset Pretrained

SoundFont-Synthesis Pretrained

Expressive-Vocal Pretrained (ours)

Soprano:

Alto:

Tenor:

Bass:

Demo #2

Mixture:

Voice

References

Non-dataset Pretrained

SoundFont-Synthesis Pretrained

Expressive-Vocal Pretrained (ours)

Soprano:

Alto:

Tenor:

Bass:

Demo #3

Mixture:

Voice

References

Non-dataset Pretrained

SoundFont-Synthesis Pretrained

Expressive-Vocal Pretrained (ours)

Soprano:

Alto:

Tenor:

Bass:

Demo #4

Mixture:

Voice

References

Non-dataset Pretrained

SoundFont-Synthesis Pretrained

Expressive-Vocal Pretrained (ours)

Soprano:

Alto:

Tenor:

Bass:

Long Separation Demos (from Cantoria Dataset and Chorale Singing Dataset):

Demo #1

Mixture:

Voice

References

Non-dataset Pretrained

SoundFont-Synthesis Pretrained

Expressive-Vocal Pretrained (ours)

Soprano:

Alto:

Tenor:

Bass: