PUB - Publikationen an der Universität Bielefeld

This paper presents a new interface for controlling sonification models. A haptic controller interface is developed which allows both to manipulate a sonification model, e.g. by interacting with it and to provide a haptic data representation. A variety of input types are supported with a hand-sized interface, including shaking, squeezing, hammering, moving, rotating and accelerating. The paper presents details on the interface under development and demonstrates application of the device for controlling a sonification model. For this purpose, the Data-Solid Sonification Model is introduced, which provides an acoustic representation of the local neighborhood relations in high-dimensional datasets for binary classification problems. The model is parameterized by a reduced data representation obtained from a growing neural gas network. Sound examples are given to demonstrate the device and the sonification model. ## Sound Demonstrations for the Audio-Haptic Ball Interface * #### Sounds for atomic collisions: Each model mass is assigned a material according to the data within the neurons voronoi cell. In our situation, binary classification data is used and the object type is either A (or B) if data from class A (or B) dominates the cell. + Collision between a A object with A object: [sound](https://pub.uni-bielefeld.de/download/2704140/2704141) (wood) + Collision between a B object and a B object: [sound](https://pub.uni-bielefeld.de/download/2704140/2704142) (plastic) + Collision between a A object and a B object: [sound](https://pub.uni-bielefeld.de/download/2704140/2704143) (glass) * #### Table 1: Sound examples for synthetic datasets from binary classification problems

File/Track:	A) two separated classes -- sound B) two classes showing little overlap along one axis  -- sound C) two classes that overlap - inseparable classes -- sound
Description:	The sonification model was excited by shaking the interface device. The shaking activation is given by the acceleration a_x(t), a_y(t) . During the first half of the sound examples, the interface is shaken along the x axis, during the second half along the y axis.
Duration:	about 5 sec.

* #### Table 2: Sound examples for Data solid sonifications using a GNG for the dataset (B) (see above) for different network complexities.

File/Track:	A: 5 neurons sound B: 10 neurons sound C: 15 neurons sound D: 20 neurons sound E: 25 neurons sound F: 30 neurons sound
Description:	Shaking excitation of the dataset with increasing network complexity.
Duration:	about 2 sec / example

* Sound Example for shaking while GNG adaptation proceeds, and thus while the data-solid structure changes over time: [sound example](https://pub.uni-bielefeld.de/download/2704140/2704153) . It can be heard that with ongoing GNG growth more and more neurons exist (more collisions). From the pitch it can be perceived that the number of data points that have a neuron as its nearest neighbor decreases - new neurons are added between neurons that are frequently activated.