Memoire Deezer | ATIAM
If audio reproduction fails, the audios are all here: Google Drive
Link
Example 1
Name: Diffusion separation output from Plaja-Roglans et al.
Trained for 99k steps.
Separated Vocals
Separated Accompaniment
Example 2
Name: Make-it-Sound-Good (MSG) bass enhancement
Trained for 50 epochs (120k steps) on the MUSDB18 train set.
Input (Spleeter's output)
Target (Ground Truth)
Output (MSG System)
Example 3
Name: Make-it-Sound-Good (MSG) accompaniment enhancement
Trained for 50 epochs (120k steps) on the MUSDB18 train set.
Input (Spleeter's output)
Target (Ground Truth)
Output (MSG System)
Example 4
Name: Soundstream by Lucidrains' output
Sample after training on 10k steps of MUSDB samples
Sample after training on 29k steps of Bean Drums
Example 5
Name: RAVE Codec Reconstruction
Comparison between the representation learning phase's output and the output after the adversarial fine-tuning phase.
Output after representation Learning Phase
The following examples first reproduce the input and then the reconstruction, of validation set audios.
Output after complete training (Adversarial Fine-tuning Phase)
The following examples first reproduce the input and then the reconstruction, of validation set audios.
Example 6
Name: Finetuneing EnCodec using embedding distance
Input (Spleeter output)
Target
Quantized Target
Output
Example 7
Name: EnCodec Super Overfit using embedding distance
Input (Spleeter output)
Output
Quantized Target
Example 8
Name: EnCodec fine-tuning for denoising using embedding distance
Input
Output
Quantized Input
Quantized Target
Example 9
Name: Residual Loss experiments
Input
Output
Quantized Input
Quantized Target
Example 10
Name: Embedding distance loss on Drums
Input
Output
Quantized Input
Quantized Target
Example 11
Name: Adding transformer layer, trained on Bass
Input
Output
Quantized Input
Quantized Target
Example 12
Name: Mix to Bass using embedding distance
Input
Output
Quantized Input
Quantized Target
Example 13
Name: Mix to Drums using embedding distance
Input
Output
Quantized Input
Quantized Target
Example 14
Name: Mix to Bass with transformer layer
Input
Output
Quantized Input
Quantized Target
Example 15
Name: Mix to Drums with transformer layer
Input
Output
Quantized Input
Quantized Target
Example 16
Name: Mix to Other using transformer layer
Input
Output
Quantized Input
Quantized Target
Example 17
Name: Mix to Bass using RAVE model
Input
Output
Quantized Input
Quantized Target
Example 18
Name: Vampnet reconstruction with Demucs as input
Input
Target
Quantized Input
Reconstruction of fine tokens from coarse (using coarse2fine model)
Reconstruction of coarse and fine tokens from 1 coarse token (using coarse and coarse2fine models)
Example 19
Name: Vampnet reconstruction with mixture as input (filtering with encoder)
Input
Target
Reconstructed Input
Reconstruction of fine tokens from coarse (using coarse2fine model)
Reconstruction of coarse and fine tokens from 1 coarse token (using coarse and coarse2fine models)
Extra examples of VampNet predictions are here:
