Acoustic echo cancellation or suppression methods aim to suppress the echo originated from acoustic coupling between loudspeakers and microphones. Conventional approaches estimate echo using adaptive filtering. Due to the nonlinearities in the acoustic path of far-end signal, further post-processing is needed to attenuate these nonlinear components. In this paper, we propose a novel architecture based on deep gated recurrent neural networks to estimate the near-end signal from the microphone signal. The proposed architecture is trained using multitask learning to learn the auxiliary task of estimating the echo in order to improve the main task of estimating the clean near-end speech signal. Experimental results show that our proposed deep learning based method outperforms the existing methods for unseen speakers in terms of the echo return loss enhancement (ERLE) for single-talk periods and the perceptual evaluation of speech quality (PESQ) score for double-talk periods.
Cite as: Fazel, A., El-Khamy, M., Lee, J. (2019) Deep Multitask Acoustic Echo Cancellation. Proc. Interspeech 2019, 4250-4254, doi: 10.21437/Interspeech.2019-2908
@inproceedings{fazel19_interspeech,
author={Amin Fazel and Mostafa El-Khamy and Jungwon Lee},
title={{Deep Multitask Acoustic Echo Cancellation}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={4250--4254},
doi={10.21437/Interspeech.2019-2908}
}