MultiDecoderDPRNN

Asteroid model

Description:

• Code: The code corresponding to this pretrained model can be found here.

• Notebook: Colab Notebook with examples can be found here

• Paper: "Multi-Decoder DPRNN: High Accuracy Source Counting and Separation", Junzhe Zhu, Raymond Yeh, Mark Hasegawa-Johnson. ICASSP(2021).

• Summary: This model achieves SOTA on the problem of source separation with an unknown number of speakers. It uses multiple decoder heads(each tackling a distinct number of speakers), in addition to a classifier head that selects which decoder head to use.

• Project Page

• Original research repo

This model was trained by Joseph Zhu using the wsj0-mix-var/Multi-Decoder-DPRNN recipe in Asteroid. It was trained on the sep_count task of the Wsj0MixVar dataset.

Training config:

filterbank:
  n_filters: 64
  kernel_size: 8
  stride: 4
masknet:
  n_srcs: [2, 3, 4, 5]
  bn_chan: 128
  hid_size: 128
  chunk_size: 128
  hop_size: 64
  n_repeats: 8
  mask_act: 'sigmoid'
  bidirectional: true
  dropout: 0
  use_mulcat: false
training:
  epochs: 200
  batch_size: 2
  num_workers: 2
  half_lr: yes
  lr_decay: yes
  early_stop: yes
  gradient_clipping: 5
optim:
  optimizer: adam
  lr: 0.001
  weight_decay: 0.00000
data:
  train_dir: "data/{}speakers/wav8k/min/tr"
  valid_dir: "data/{}speakers/wav8k/min/cv"
  task: sep_count
  sample_rate: 8000
  seglen: 4.0
  minlen: 2.0
loss:
  lambda: 0.05

Results:

'Accuracy': 0.9723333333333334, 'P-Si-SNR': 10.36027378628496

License notice:

This work "MultiDecoderDPRNN" is a derivative of CSR-I (WSJ0) Complete by LDC, used under LDC User Agreement for Non-Members (Research only). "MultiDecoderDPRNN" is licensed under Attribution-ShareAlike 3.0 Unported by Joseph Zhu.