Downloads · 30 days
0
DPTNet_jaCappella_VES_48k
jaCappella/DPTNet_jaCappella_VES_48k
DPTNet_jaCappella_VES_48k is a audio-to-audio model from jaCappella. Use it for the audio-to-audio task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-nc-4.0.
This model was trained by Tomohiko Nakamura using the codebase). It was trained on the vocal ensemble separation task of the jaCappella dataset. The paper was published in ICASSP 2023 (arXiv).
Access
Public
Updated Feb 12, 2024
Repo size
45.6 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.pth45.6 MB · 100%
At a glance
- Task
- Audio-to-Audio
- License
- cc-by-nc-4.0
- Access
- Public
- Created
- Jan 21, 2023
- Updated
- Feb 12, 2024
- SHA
- 00e6e620
Datasets
Model card
From the Hugging Face model README
DPTNet trained with the jaCappella corpus for vocal ensemble separation
This model was trained by Tomohiko Nakamura using the codebase).
It was trained on the vocal ensemble separation task of the jaCappella dataset.
The paper was published in ICASSP 2023 (arXiv).
License
See the jaCappella dataset page.
Citation
See the jaCappella dataset page.
Configuration
data:
num_workers: 12
sample_rate: 48000
samples_per_track: 13
seed: 42
seq_dur: 5.046
source_augmentations:
- gain
sources:
- vocal_percussion
- bass
- alto
- tenor
- soprano
- lead_vocal
filterbank:
kernel_size: 32
n_filters: 64
stride: 16
masknet:
bidirectional: true
chunk_size: 174
dropout: 0
ff_activation: relu
ff_hid: 256
hop_size: 128
in_chan: 64
mask_act: sigmoid
n_repeats: 8
n_src: 6
norm_type: gLN
out_chan: 64
optim:
lr: 0.005
optimizer: adam
weight_decay: 1.0e-05
training:
batch_size: 1
early_stop: true
epochs: 600
gradient_clipping: 5
half_lr: true
loss_func: pit_sisdr
Results (SI-SDR [dB]) on vocal ensemble separation
| Method | Lead vocal | Soprano | Alto | Tenor | Bass | Vocal percussion |
|---|---|---|---|---|---|---|
| DPTNet | 8.9 | 8.5 | 11.9 | 14.9 | 19.7 | 21.9 |
- Task
- Audio-to-Audio
- License
- cc-by-nc-4.0
- Languages
- ja
- Created
- Jan 21, 2023
- Updated
- Feb 12, 2024