Asteroid model cankeles/ConvTasNet_WHAMR_enhsingle_16k
Description:
This model was fine tuned on a modified version of WHAMR! where the speakers were taken from audiobook recordings and reverb was added by Pedalboard, Spotify.
The initial model was taken from here: https://huggingface.co/JorisCos/ConvTasNet_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:
'sar': 13.612368475881558,
'sar_imp': 9.709316571584433,
'sdr': 13.612368475881558,
'sdr_imp': 9.709316571584433,
'si_sdr': 12.978640274976373,
'si_sdr_imp': 9.161273840297232,
'sir': inf,
'sir_imp': nan,
'stoi': 0.9214516928197306,
'stoi_imp': 0.11657488247668318