Object-oriented implementation of Auditory Front-End

Contents

This Demo scripts presents the object-oriented implementation of the Two!Ears Auditory Front-End. It is intended as a demo at that stage, and not a working copy, as important feature are not implemented fully yet.

Currently supported features

getDependencies('available')

ans = 

  Columns 1 through 5

    'time'    'filterbank'    'innerhaircell'    'adaptation'    'ams_features'

  Columns 6 through 10

    'crosscorrelation'    'autocorrelation'    'ratemap'    'ild'    'itd'

  Columns 11 through 14

    'ic'    'spectral_features'    'onset_strength'    'offset_strength'

  Columns 15 through 18

    'pitch'    'onset_map'    'offset_map'    'gabor'

Coming-up features:

Initialize and load a signal

clear all,close all
load('TestBinauralCues');

1- Basic advantages of object-oriented approach: Extraction of inner hair-cell envelope for a single audio channel

request = 'innerhaircell';

% Select one channel of the audio signal
data = earSignals(:,2);

Initialization is done in two steps:

dObj = dataObject(data,fsHz); % Create a data object based on this signal
mObj = manager(dObj,request); % Instantiate a manager

Everything is now ready for processing:

mObj.processSignal

Results are stored in the data object:

dObj.innerhaircell{1}.plot;

The object dObj contains all the signals computed in the process as objects, each having specific methods, e.g.:

fprintf('%s for %s signal was computed for %i frequency channels.\n',...
    dObj.innerhaircell{1}.Label,...
    dObj.innerhaircell{1}.Channel,...
    size(dObj.innerhaircell{1}.cfHz,2))

Inner hair-cell envelope for mono signal was computed for 31 frequency channels.

Most objects are inheriting from the "handle" super-class, which allows to call them "by reference" instead of Matlab typical "by value", limiting memory usage from duplicates and allowing for using pointers to objects instances instead of duplicates. For example, the input and output lists properties of the manager class (used for routing inputs and outputs) are pointers to the actual input/output signals in the data object:

mObj.OutputList{1}.plot;
mObj.InputList{1}.play

clear dObj mObj data

2- Complex scenario:

Chunk-based extraction of a frame-based binaural cue (ILD)

request = 'ild';
chunkSize = 2000;   % Chunk duration in samples

% Use both ear signals
data = earSignals;

Create an empty data object. It will be filled up as new ear signal chunks are "acquired".

dObj = dataObject([],fsHz,10,2);   % Last input (2) indicates a stereo signal
mObj = manager(dObj,request);   % Instantiate a manager

From here on, simulating real-time chunk acquisition and processing request...

% Number of chunks in the signal
n_chunks = ceil(size(data,1)/chunkSize);

% Zero-pad the signal to an integer number of chunks (not necessary, but
% but used for later comparison with signal-based processing)
data = [data;zeros(n_chunks*chunkSize-size(data,1),size(data,2))];

% Loop on all the chunks
for ii = 1:n_chunks

Each new chunk of signal is then read and processed by the manager

    % Read signal chunk
    chunk = data((ii-1)*chunkSize+1:ii*chunkSize,:);

    % Request processing for that chunk
    mObj.processChunk(chunk,1);

end

% Plot the result
dObj.ild{1}.plot;

Extract the same cue but from the whole signal at once for comparison

dObj2 = dataObject(data,fsHz);  % Create a new data object
mObj2 = manager(dObj2,request); % Create a new manager
mObj2.processSignal;            % Do the processing

Compute and plot the difference between the two approaches

delta_ild = TimeFrequencySignal(dObj.ild{1}.FsHz,dObj.bufferSize_s,'ild',dObj.ild{1}.cfHz,...
    'Difference in ILD, chunk vs. signal',...
    dObj.ild{1}.Data(:)-dObj2.ild{1}.Data(:));
delta_ild.plot;

Published with MATLAB® R2014a