Two channels in audio recording, how to perform analysis?

This is a rather simple question of audio processing basics that I am struggling to understand. When analyzing audio in MATLAB or any program, what are you supposed to do with multi-channel audio? Do you ignore one channel, like in the example I have below, and only look at the first channel? I understand that the 2 channels represent the left and the right channels, just confused why you can ignore 1 channel when doing fft, plotting spectrograms, etc. Are they the same? Do I lose information if I record to 1 channel instead of 2?

I am working on a project using machine learning to identify anomolies in audio data, and am trying to make sure I am collecting/analyzing data in the right manner.

fname = 'example.wav';

info = audioinfo(fname);

[y,Fs] = audioread(fname);

y=y(:,1);

Answer 1

[Image Analyst]

It depends on what you want to do. Perhaps anomolies can be identified in both channels, like a thunderbolt. But if you want to know what side of the "microphone" a dog suddenly barks on? Then you'd need both channels.

What kind of data do you have? Do you have any examples of an audio file that has the kind of "anomalies" that you're trying to find? I mean it could be anything from a pop in an old vinyl record, to a single man talking in a crowd of talking women. So there is an infinite number of things that could be defined as "unusual", depending on the context.

https://en.wikipedia.org/wiki/Auditory_illusion

Have you tried looking at the spectrogram to find places where the spectrum suddenly changes? Try the spectrogram() function in the Signal Processing Toolbox.

[Ivan Arkhipov]

I see what you mean with direction potentially being important. I think in my case it likely is not necessary to know the direction but rather to identify a sharp change in the audio signal, which would likely continue for an extended period of time. The project is involved with identifying anomalies in the cold spray thermal spray process, used to deposit metal powders on materials. There are many possible sources of anomilies in this process (nozzle clogging, changing powder feed rate, and other parameter variations). I do not have a lot of data at the moment, as I just bought a high-end microphone. I plan to collect a lot of data soon by introducing different anomalies, varying different parameters, spraying on different materials, etc.

If you're interested, I attached two audio files I had been working on from an initial experiment conducted at two different gas temperatures with four incremental changes to the gas pressure. These pressure changes are clearly identifiable in the audio signal and the spectrogram, but the temperature difference is not clear. This could just be due to temperature change not having a significant effect on acoustics.

I actually made a simple image classification CNN to try to classify validation data into one of two temperatures and one of four pressures. I did this by splitting each audio segment (at different pressures) into 20 segments, each being a second long. I used spectrograms from 18 of these segments at each temperature and pressure as training data and left the other 2 from each segmentation as validation data. This CNN was effective in classifying each validation file to the right pressure but not the right temperature. I did all of this analysis just using one channel of audio, as oppose to both, so that was where my question stemmed from. Is there a way to make a spectrogram from multiple channels?

Sorry if that was all too much information.

[Image Analyst]

Where exactly do you think something unusual is happening?

% Demo by Image Analyst
clc;    % Clear the command window.
close all;  % Close all figures (except those of imtool.)
clear;  % Erase all existing variables. Or clearvars if you want.
workspace;  % Make sure the workspace panel is showing.
format long g;
format compact;
fontSize = 18;
fprintf('Beginning to run %s.m ...\n', mfilename);

fileList = dir('*.mp3')
fullFileName = fullfile(fileList(1).folder, fileList(1).name);
info = audioinfo(fullFileName)
[y,Fs] = audioread(fullFileName);
y1 = y(:,1);
y2 = y(:,2);
subplot(2, 1, 1);
plot(y1, 'b-', 'LineWidth', 1);
grid on;
title('Left Channel', 'FontSize', fontSize);
xlabel('Index', 'FontSize', fontSize);
ylabel('Amplitude', 'FontSize', fontSize);

subplot(2, 1, 2);
plot(y2, 'b-', 'LineWidth', 1);
grid on;
title('Left Channel', 'FontSize', fontSize);
xlabel('Index', 'FontSize', fontSize);
ylabel('Amplitude', 'FontSize', fontSize);
g = gcf;
g.WindowState = 'maximized';

figure
spectrogram(y1);

fprintf('Done running %s.m.\n', mfilename);

I haven't listened to the waveforms but from the looks of them it looks like there are two boosts in volume, at about 25% and 50% of the way through.

[Ivan Arkhipov]

There isn't necessarily anything unusual happening here. It is the beginning of a long term study and this was the first experiment conducted. The pressure of the gas stream was changed manually at the 3 "boosts in volume" like you said at 25%, 50% and 75%. If the pressure changed unexpectedly, it is clear that such an anomolous event could be detected with audio based on these files. The two files are also recorded at different temperatures, so it could be said that the second file is an anomolous version of the first file (i.e 400C vs 300C). It does not seem like it is possible to distinguish the two, so I was curious if looking at both channels or any other analysis would be effective.

[Image Analyst]

Why not just see when the amplitude peaks leave some control limits? Or else see if the signal now is much different than it was 5 or 10 seconds ago?

[Ivan Arkhipov]

That is along the lines of what I intend to do, for general anomaly detection. Pressure is just one parameter, though, and there may not be such a clear volume difference with others. Also, this type of analysis would still not allow me to distinguish the two files at different temperatures, or tell if the temperature is changing in a new recording. It does not seem like there is noticeable difference between the two audios here, or atleast not with the analysis methods I have tried (looking at power spectrum and spectrograms).

[Image Analyst]

The solution then may be a deep learning network where you train it with waveforms that are labeled "normal" or "abnormal".