metadata
tags:
- asteroid
- audio
- ConvTasNet
- audio-to-audio
datasets:
- WHAMRmod
- enh_single
license: cc-by-sa-4.0
Asteroid model cankeles/ConvTasNet_WHAMRmod_enhsingle_16k
Description:
This model was fine tuned on a modified version of WHAMR! where the speakers were taken from audiobook recordings.
The initial model was taken from here: https://huggingface.co/JorisCos/DCCRNet_Libri1Mix_enhsingle_16k
This model was trained by M. Can Keles using the WHAM recipe in Asteroid.
It was trained on the enh_single task of the WHAM dataset.
Training config:
data:
mode: min
nondefault_nsrc: null
sample_rate: 16000
task: enh_single
train_dir: wav16k/min/tr/
valid_dir: wav16k/min/cv/
filterbank:
kernel_size: 16
n_filters: 512
stride: 8
main_args:
exp_dir: exp/tmp
help: null
masknet:
bn_chan: 128
hid_chan: 512
mask_act: relu
n_blocks: 8
n_repeats: 3
n_src: 1
skip_chan: 128
optim:
lr: 0.001
optimizer: adam
weight_decay: 0.0
positional arguments: {}
training:
batch_size: 2
early_stop: true
epochs: 10
half_lr: true
num_workers: 4
Results:
On Libri1Mix min test set :
si_sdr: 14.743051006476085
si_sdr_imp: 11.293269700616385
sdr: 15.300522933671061
sdr_imp: 11.797860134458015
sir: Infinity
sir_imp: NaN
sar: 15.300522933671061
sar_imp: 11.797860134458015
stoi: 0.9310514162434267
stoi_imp: 0.13513159270288563