In this thesis, we propose an acoustic echo cancellation (AEC) algorithm based on
a stacked bidirectional long short-term memory (BLSTM). Echo is caused by the sound produced by a loudspeaker being picked up by a microphone in the same room. AEC is a ...
In this thesis, we propose an acoustic echo cancellation (AEC) algorithm based on
a stacked bidirectional long short-term memory (BLSTM). Echo is caused by the sound produced by a loudspeaker being picked up by a microphone in the same room. AEC is a device for removing echo and transmitting clear near-end speech through the channel to the far-end.
Linear echo canceller implemented using an adaptive filter often produces speech
distortions, depending on the background noise and communication environment. Sound quality is also degraded in a double-talk situation where both far-end and near-end speakers talk simultaneously.
To solve this problem, the deep neural network (DNN) technology, which has been successfully used in various fields, is applied to AEC. In this thesis, BLSTM, a kind of recurrent neural network (RNN), is stacked to separate and remove echo and background noise from the microphone signal. To this end, we propose a new BLSTM stacking technique, and an attention mechanism is introduced to BLSTM cells to improve the speech separation performance.
Experimental results show that the proposed BLSTM stacking method has better
speech separation performance than the conventional method, and the proposed DNN based AEC algorithm is superior in many ways to the conventional AEC based on the adaptive filter.