• TasNet - Time-domain Audio Separation Network
  • Architecture
  • Experiment
  • Reference

Watch the NTU cource by youtube.

TasNet - Time-domain Audio Separation Network

Architecture

• Encoder

  • 512-d : 不一定要positive

  

• Decoder

  • 跟Encoder互為Inverse的效果沒比較好

  

• Separator

  • Network Compression
    • Depthwise Separable Convolution
      • Steps
        1. Depthwise Convolution
           • Filter number = Input channel number
           • Each filter only consider on channel
           • The filters are k x k matrices
           • There is no interaction between channels
        2. Pointwise Convolution
           • 1 x 1 filter
      • Application
        • SqueezNet
        • MobileNet
        • ShuffleNet
        • Xception

  

Experiment

Reference

• Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation, Y Luo, 2019
• 李宏毅-Speech Separation (2/2) - TasNet
• Speech Separation-李宏毅 HUNG-YI LEE
• 李宏毅-Network Compression (5/6)