Data handling

Circular buffer

Memory pre-allocation of large arrays in Matlab is well known to be a critical operation for optimising computation time. The Auditory front-end, particularly in an online scenario, will be confronted with this problem. For each new chunk of the input signal, chunks of output are computed for each internal representation and are appended to the already existing output. Computation time will be strongly affected if the arrays containing the data are not initialised appropriately (i.e., the memory it occupies is pre-allocated) to fit the input signal duration.

The issue in a real-time scenario is that the signal duration is unknown. To overcome this problem, data for each signal is stored in a buffer of fixed duration which is itself pre-allocated. Buffers are updated following a FIFO rule: once the buffer is full, the oldest samples in the buffer are overwritten by the new signal samples.

The circVBuf class

A conceptual way of implementing a FIFO rule is to use circular (or ring) buffers. The inconvenience of a traditional linear buffer is that once it is full and new input overwrites old samples (i.e., it is in its “steady-state”), reading the data from it implies reaching the end of the buffer and continuing reading from its beginning. The data read will be in two fragments, because of the linear buffer having a physical beginning and end which do not match to the oldest and newest data samples. This is eliminated in circular buffers which do not have a beginning or end, and a contiguous segment is always obtained upon reading. Circular buffers were implemented for the Auditory front-end based on the third-party class provided by [Goebbert2014], which has been slightly modified to account for multi-dimensional data (instead of vector-only).

Circular buffer interface

The circVBuf class provides a buffer that is conceptually circular, in the sense that it allows continuous reading of the data. However in practice it still stores data in a linear array in Matlab (the size of which is, however, twice the size of the actual data). Accessing stored data requires knowledge about this class and can be tedious to a naive user. To eliminate confusion and make the buffer transparent to the user, the interface circVBuffArrayInterface was implemented, with the aim of allowing the buffer to use most basic array operations.

Given a circular buffer circBuffer, the interface is obtained by

buffer = circVBufArrayInterface(circBuffer)

It will allow the following operations:

  • buffer(n1:n2) returns stored data between positions n1 and n2, where position 1 is the oldest sample in the buffer (but not necessarily the first one in the actual array storing data, due to circularity). For multiple dimensions, these indices always refer to the first dimension. To return stored data up to the most recent sample, use buffer(n1:end).
  • buffer(:) returns all data stored in the buffer (ignoring “empty” sections of the buffer, if said buffer was never filled).
  • buffer(’new’) returns the latest chunk of data that was added to the buffer.
  • length(buffer) returns the effective (i.e., ignoring empty sections) buffer length across its first dimension.
  • size(buffer) returns the effective size of the buffer (including other dimensions).
  • numel(buffer) returns the total number of elements stored (calculated as product of the effective dimensions).
  • isempty(buffer) returns true when no data is stored, false otherwise.

This provides an array behaviour to the buffers, simplifying greatly their use.

Note

Note that the only limitation is the need of the column operator : to access all data, as in buffer(:). Without it, buffer will return a handle to the circVBufArrayInterface object.

Signal objects

Signals are implemented as objects in the Auditory front-end. To avoid code repetition and make better use of object-oriented concepts, signals are grouped according to their dimensions, as they then share the same properties. The following classes are implemented:

  • TimeDomainSignal for one-dimensional (time) signals.
  • TimeFrequencySignal which stores two-dimensional signals where the first dimension relates to time (but can be, e.g., a frame index) and the second to the frequency channel. These signals include as an additional property a vector of channel centre frequencies cfHz. Signals of such form are obtained from requesting, for example, ’filterbank’, ’innerhaircell’,’ild’,... In addition, time-frequency signals containing binary data (used e.g., in onset or offset mapping) have their own BinaryMask signal class.
  • CorrelationSignal for three-dimensional signals where the third dimension is a lag position. These include also the cfHz property as well as a vector of lags (lags).
  • ModulationSignal for three-dimensional signals where the third dimension is a modulation frequency. These include cfHz and modCfHz (vector of centre modulation frequencies) as properties.
  • FeatureSignal used to store a collection of time-domain signals, each associated to a specific name. Each feature is a single vector, and all of them are arranged as columns of a same matrix. Hence they include an ordered list of features names fList that labels each column.

All these classes inherit the parent Signal class. Hence they all share the following common “read-only” properties:

  • Label, which is a “formal” description of the signal, e.g., ’Inner hair-cell envelope’, used for example when plotting the signal.
  • Name, which is a name tag unique to each signal type, e.g., ’innerhaircell’. This name corresponds to the name used for a request to the manager.
  • Dimensions, which describes in a short string how dimensions are arranged in the signal, e.g., ’nSamples x nFilters’
  • FsHz, the sampling frequency of this specific signal. If the signal is framed or down-sampled (e.g., like a rate-map or an ILD) this value will be different from the input signal’s sampling frequency.
  • Channel, which states ’left’, ’right’ or ’mono’, depending on which channel from the input signal this signal was derived.
  • Data, an interface object (circVBufArrayInterface described earlier) to the circular buffer containing all data. The actual buffer, Buf is a circVBuf object and a protected property of the signal (not visible to the user).

The Signal class defines the following methods that are then shared among children objects:

  • A super constructor, which sets up the internal buffer according to the signal dimensions. Each children signal class is calling this super constructor before populating its other properties.
  • An appendChunk method used to fill the internal buffer.
  • A setData method used for initialising the internal buffer given some data.
  • A clearData method for re-initialisation.
  • The getSignalBlock method returning a segment of data of chosen duration, starting from the newest elements.
  • The findProcessor method which, given a handle to a manager object, will retrieve which processor has computed this specific signal (by comparing it with the Output property of each processor, described in General considerations).
  • A getParameters method which, given a handle to a manager object, will retrieve the list of parameters used in the processing to obtain that signal.

In addition, the Signal class defines an abstract plot method, which each children should implement. This cannot be defined in the parent class as the plotting routines will be drastically different depending on children signal dimensions. Children classes therefore only implement their own constructor (which still calls the super-constructor) and their respective plotting routines.

Data objects

Description

Many signal objects are instantiated by the Auditory front-end (one per representation involved and per channel). To organise and keep track of them, they are collected in a dataObject class. This class inherits the dynamicprops Matlab class (itself inheriting the handle) class. This allows to dynamically define properties of the class.

This way, each signal involved in a given session of the Auditory front-end will be grouped according to its class in a distinct property of the dataObject, with name given by the signal signal.Name unique name tag. Extra properties of the data object include:

  • bufferSize_s which is the common duration of all circVBuf objects in the signals.
  • A flag isStereo, which if true will indicate to the data object that all signals come as pairs of left/right channels.

Data objects are constructed by providing an input signal (which can be empty in online scenarios), a mandatory sampling frequency in Hz, a global buffer size (10 s by default), and the number of channels of the input (1 or 2). This number of channel is not necessary if an input signal is used as argument in the constructor but needs to be provided otherwise.

The dataObject definition includes the following, self-explanatory methods:

  • addSignal(signalToAdd)
  • clearData
  • getParameterSummary returning a list of all parameters used for the computation of all included signal (given a handle to the corresponding manager).
  • play, provided for user convenience.

Signal organisation

As mentioned before, data objects store signal objects. Each class of signal occupies a property in the data object named after the signal .Name property. Multiple signals of the same class will be stored as a cell array in that property. In the cell array, the first column is always for the left channel (or mono signal), and the second column for the right channel. If multiple signals of the same type are present (e.g., if the user requested the same representation twice but with a change of parameters), then the corresponding signals are stored in different lines of the array. For instance, for a session where the user requested the inner hair-cell envelope twice, with the second request changing only the way of extracting the envelope (i.e., the parameter ’ihc_method’), the following data object is created:

>> dataObj

dataObj =

  dataObject with properties:

     bufferSize_s: 10
         isStereo: 1
             time: {[1x1 TimeDomainSignal]  [1x1 TimeDomainSignal]}
            input: {[1x1 TimeDomainSignal]  [1x1 TimeDomainSignal]}
        gammatone: {[1x1 TimeFrequencySignal]  [1x1 TimeFrequencySignal]}
    innerhaircell: {2x2 cell}

Each signal-related field except innerhaircell is a cell array of a single line (one signal), and two columns (for left and right channel). Because the second request from the user included only a change in parameter for the inner hair-cell computation, the same initial gammatone signal is used for both, but there are two output innerhaircell signals (hence a cell array of two lines) for each channel (hence two columns).

In that case, to distinguish between the two signals and know which one was computed with which set of parameter, one can call the signal’s getParameters method. Given a handle to the manager object, it will return a list of all parameters used to obtain that signal (including parameters used in intermediate processing steps).