92 M. Gollnick et al. Fig. 9.2 Left: Moving peak from 1 kHz to 9 kHz. Right: Behavior of moving peaks in the feature space Fig. 9.3 Left: Example of an indicator function with one peak to show the relative height. Right: Indicator function transformed to feature space current frequency band between [0..1]. The resulting feature is independent of different altitude levels in indicator functions. In summary, a peak is defined either by a high left and right prominence and/or by a high relative height value. 9.4 Generation of Training Data The training data can consist of real measurement data or synthetic data. To insert measurement data, it must be manually pre-processed so that peaks and non-peaks are marked. This process requires a considerable amount of additional work, since a large amount of data must be available in order to map sufficient variations in peaks and each data set must also be processed manually. In addition, errors can always occur during preprocessing by humans, such as incorrectly marked peaks, which have a negative impact on the classification. Therefore, synthetic training data were used in this paper. Thus, expertise on peaks and non-peaks contributes to the structure of synthetic indicator functions and it is possible to correct and/or extend the training set at any time.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==