[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
auditory models and self-organising maps
thank you very much for your insightful replies (Peter, Martin & Jont) to my
inquiry concerning the frame rate for auditory models -- they were really
helpful, but raised of course new questions as well.
Jim Stevenson asked me to tell more about the self-organizing map, therefore
I post a more detailed description of this stage:
For different genres of music it is generally not possible to arrange the
multitude of different sound events in pre-defined classes, as it can be
done with phonemes for speech. It is therefore tempting to use a
self-organising articial neural network for the classifation of the
pre-processed signals, such as he self-organising map (SOM). It has been
developped by Teuvo Kohonen and was inspired by feature maps in the cerebral
cortex. A SOM is able
to map high-dimensional input signals on a two- or three-dimensional grid
while preserving their topological relation so that similar input signals
are usually mapped next to each other. The feature map thus provides a
measure of similarity. In this case the input signals are constituted by a
sequence of vectors from the auditory model, corresponding to a sound event.
Each vector is mapped to a point on the feature map, and the whole sequence
be represented as a trajectory. Like all artifcial neural networks, a
self-organising map needs to be trained to adapt its weight vectors to the
distribution of input signals. The training data is presented to the network
up to 100,000 times during the ordering process. The spectrum of training
data should be similar to the data intended for classification to achieve
the best results, although a SOM is able to generalise to a certain extent.
Our goal is to largely match the classification results of the SOM with our
perception, although it is not easy to define what 'perceptual simililarity'
means for dynamic sounds.
Thank you for any further suggestions,
University of Hertfordshire, UK