Multi-Channel Target Speech Extraction with Channel Decorrelation and Target Speaker Adaptation

The codes here are implementations of two methods for exploiting the multi-channel spatial information to extract the target speech. The first one is using a target speech adaptation layer in a parallel encoder architecture. The second one is designing a channel decorrelation mechanism to extract the inter-channel differential information to enhance the multi-channel encoder representation.

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

Data

• 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.

Usage

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

Results

System	IPD	Adap	SDR/SiSDR
(0) TD-SpkBeam	-	-	11.51 / 11.00
(1)	Y	-	11.57 / 11.07
(2) Parallel	-	-	12.43 / 11.91
(3)	-	Y	12.73 / 12.20
(4) CD	-	-	12.87 / 12.34
(5)	-	Y	12.87 / 12.35
(6)	Y	Y	12.55 / 12.01
(7) CC	-	Y	12.66 / 12.13

Contact

Email: [email protected]

Reference

• 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.