search

ldzhangyx / MusicMagus

The official implementation of the IJCAI 2024 paper "MusicMagus: Zero-Shot Text-to-Music Editing via Diffusion Models".
Apache License 2.0
26 stars 0 forks source link

Experimental dataset #2

Open XinleiNIU opened 1 month ago

XinleiNIU commented 1 month ago

Hi, I found your work is interesting, would it be possible to release the data used for the timbre transfer task? Thank you in advance.

ldzhangyx commented 1 month ago

You mean the evaluation dataset?

XinleiNIU commented 1 month ago

You mean the evaluation dataset?

Hi Yixiao,

Yes, the audio samples used in the objective experiments.

ldzhangyx commented 1 month ago

Hi Xinlei,

Sorry for the late reply. Since the audio samples in the objective experiments are randomly generated, I did not store them after leaving Sony. To reproduce the data pairs, you can perform 2 inferences with and without cross-attention guidance. The model should perform consistently.

XinleiNIU commented 1 month ago

Hi Yixiao,

Thanks for the reply. Can the model transfer the same composition into several different musical instruments? The reconstruction quality is not good for the audio downloaded from the demo page. Do you have any suggestions for tuning the model? Ideally, I'd like to obtain paired musical composition with different timbre performance. Thank you very much!