Continuous Audio Object Recognition Diploma Thesis

The detection of sound events is a key technology for a various set of audio applications. Sounds are able to transport information through vision borders. Therefore, a humanoid robot assigned with kitchen tasks improves its interactive behavior with the environment a lot when using acoustics. While audio scene analysis employs a lot of subjects, this thesis deals with the recognition of presegmented as well as continuous audio objects using single channel microphone input. Further prior knowledge on scenes with single and multiple sources was not used. This means that recognition is performed without information on the audio context like source positions and statistical information on typical event sequences. The three explored feature sets consisted of MFCCs with first and second order temporal derivatives, PCA-ICA features without using temporal context and PCA-ICA features on several context window sizes. In a first batch of experiments those features were evaluated for GMMs, forward and ergodic HMMs on predefined segments for single source data, which was recorded in different kitchens. The results show that for single source data MFCC features perform worse than ICA features, independent of the classifier. Further, ICA features covering temporal context gave even better results. The comparison of forward and ergodic models for different number of states revealed that the kitchen task class set generally favors ergodic HMMs instead of left-right models. Another experiment confirmed the superiority of ICA to MFCCs with respect to the number of gaussian parameters. While the ICA features for an architecture, which cover shared global interclass properties, appeared to be superior on single source data, this benefit could not be shown under continuous real world cooking conditions with background noise. Scarce class occurrences in realworld conditioned data in combination with low recognition performance showed that source separation, confidence measures and multi track hypothesis output need to be considered in future research directions. Furthermore, the mapping of acoustic entities to semantics during labeling and training has to be performed carefully.