Asteroid模型 JorisCos/ConvTasNet_Libri3Mix_sepnoisy_16k

描述:

这个模型是由Joris Cosentino使用librimix配方在Libri3Mix数据集的sep_noisy任务上进行训练的。

训练配置:

data:
  n_src: 3
  sample_rate: 16000
  segment: 3
  task: sep_noisy
  train_dir: data/wav16k/min/train-360
  valid_dir: data/wav16k/min/dev
filterbank:
  kernel_size: 32
  n_filters: 512
  stride: 16
masknet:
  bn_chan: 128
  hid_chan: 512
  mask_act: relu
  n_blocks: 8
  n_repeats: 3
  n_src: 3
  skip_chan: 128
optim:
  lr: 0.001
  optimizer: adam
  weight_decay: 0.0
training:
  batch_size: 8
  early_stop: true
  epochs: 200
  half_lr: true
  num_workers: 4

结果:

在Libri3Mix最小测试集上:

si_sdr: 5.926151147554517
si_sdr_imp: 10.282912158535625
sdr: 6.700975236867358
sdr_imp: 10.882972447337504
sir: 15.364110064569388
sir_imp: 18.574476587171688
sar: 7.918866830474568
sar_imp: -0.9638973409971135
stoi: 0.7713777027310713
stoi_imp: 0.2078696167973911

许可声明:

这个作品"ConvTasNet_Libri3Mix_sepnoisy_16k"是Vassil Panayotov的 LibriSpeech ASR corpus 的衍生作品，并使用 CC BY 4.0 下的The WSJ0 Hipster Ambient Mixtures数据集；"ConvTasNet_Libri3Mix_sepnoisy_16k"由Joris Cosentino根据 Attribution-ShareAlike 3.0 Unported 许可