popcornell/DeMask_Surgical_mask_speech_enhancement_v1

Description:

This model was trained by popcornell using the asteroid/demask recipe in Asteroid. It was trained on the enhancement task of the Surgical_mask_speech_enhancement_v1 dataset.

Training config:

• positional arguments:
• filterbank:
  • fb_type: stft
  • n_filters: 512
  • kernel_size: 512
  • stride: 256
• demask_net:
  • input_type: mag
  • output_type: mag
  • hidden_dims: [1024]
  • dropout: 0
  • activation: relu
  • mask_act: relu
  • norm_type: gLN
• data:
  • fs: 16000
  • length: 4
• optim:
  • lr: 0.001
  • weight_decay: 1e-05
• training:
  • epochs: 200
  • batch_size: 8
  • gradient_clipping: 5
  • accumulate_batches: 1
  • save_top_k: 10
  • num_workers: 8
  • patience: 30
  • half_lr: True
  • early_stop: True
  • gaussian_mask_noise_snr_dB: np.random.randint(3, 12)
  • white_noise_dB: np.random.randint(-3, 30)
  • speed_augm: np.random.uniform(0.95, 1.05)
  • gain_augm: np.random.randint(-30, -2)
  • n_taps: 97
• main_args:
  • help: None
  • exp_dir: exp/tmp

Results:

License notice:

This work "DeMask_Surgical_mask_speech_enhancement_v1_enhancement" is a derivative of LibriSpeech ASR corpus by Vassil Panayotov, used under CC BY 4.0; of Free Universal Sound Separation Dataset by Scott Wisdom; Hakan Erdogan; Dan Ellis and John R. Hershey, used under Creative Commons Attribution 4.0 International. "DeMask_Surgical_mask_speech_enhancement_v1_enhancement" is licensed under Attribution-ShareAlike 3.0 Unported by popcornell.