The codes here are implementations of attention-based scaling adaptation (ASA), which introduces a special attention mechanism to investigate the dynamic interaction between different mixtures and the target speaker to exploit the discriminative target speaker clues. Under the same network configurations, the ASA in a single-channel condition can achieve competitive performance gains as that achieved from two-channel mixtures with inter-microphone phase difference (IPD) features.

To optimize the memory during training process, the memonger technology is applied.

• Mixtures, cd ./data/anechoic, modify corresponding data path and run sh run_data_generation.sh.
• Reverberation, cd ./data/reverberate, modify corresponding data path and run sh launch_spatialize.sh.

• Training: configure the conf.py and run sh train.sh 0 asa.
• Evaluate: modify the corresponding data of decode.sh and run sh decode.sh asa.

Email: [email protected]

• Code & Scripts
  • Mixture data
  • Reverberate data
  • Conv-TasNet implementation
• Paper
  • M. Delcroix, T. Ochiai, K. Zmolikova, K. Kinoshita, N. Tawara, T.Nakatani, and S. Araki, “Improving speaker discrimination of target speech extraction with time-domain speakerbeam,” in Proc. ICASSP. IEEE, 2020, pp. 691–695.