Using convolutional neural network in acoustic echo cancellation

Hello, i'm trying to code a neural network (NN) that can convolve a signal with a room impulse respons (RIR) to obtain an echo signal as a proof of concept. I'm using the TIMIT database and I've pre-processed it so that all signals contained within are the same length by adding zeros to the end of the signals until they match the largest signal.

So far I've used a simple NN of 5 layers and less, going from just a convolution layer in between an image input layer and a regression output layer to including a batchNorm and a reluLayer. I can't use more than one convolution layer or the signals are reduced to zero. In terms of results, with just the convolution layer the signal doesn't appear to change beyond reducing its magnitude immensely. I can barely hear it even with headphones when I reproduce with the sound function. If I include more layers, the signal is distorted, ending more as a noisy signal rather than an echo signal.

Could someone help me, please?

Answer 1

[Harsh]

Hi @José Paiva

Assuming that by “coding a neural network to convolve a signal with a room impulse response (RIR),” you want to train a 1D CNN to replicate the echo effect of linear convolution. Here's how you can do this:

1. Simulate Ground Truth in Simulink

• Use "From Workspace" to load your audio and RIR.
• Use "FIR Filter" to convolve the signal with the RIR.
• Use "To Audio Device" or "Time Scope" to listen or visualize the echoed signal.

2. Design a Minimal 1D CNN in MATLAB

• Use "sequenceInputLayer" to input the raw signal.
• Use "convolution1dLayer" with Padding="same" and filter size equal to RIR length.
• Use "regressionLayer" to output the convolved signal.

3. Train and Deplo the Network

• Use "trainnet" with original signals as input and FIR-filtered signals as targets.
• Load the trained network using the "Network" block.
• Compare its output with the FIR result using "Time Scope" or "Spectrum Analyzer".

Run the following commands in MATLAB command window to get respective documentation for more information-

doc firfilter
doc sequenceInputLayer
doc convolution1dLayer
doc trainnet