Self-learning and intelligently acting machines have the potential to be valuable and helpful companions in everyday life. The concept of learning includes the discovery of recurring patterns in given data. A prominent example is the processing of recorded microphone signals.
The objective of Unsupervised Learning is the discovery of unknown structure without knowledge of its presence and composition. In the context of our research, we are especially enganged in the discovery of recurring patterns in speech. This includes, for instance, words, syllables and phonemes and can thus be summarized as the Unsupervised Learning of structure in speech. Another application of structure discovery in audio data is the detection of acoustic events like knocking, laughter and clapping.
The research is taking place in the context of the DFG Priority Programme "Autonomous Learning" and in cooperation with the Machine Learning for Signal Processing Group of the Carnegie Mellon University, Pittsburgh, USA.
- Unsupervised Segmentation of Letter and Phoneme Sequences or Letter and Phoneme Lattices
- Unsupervised Discovery of Acoustic Units for Automatic Speech Acquisition
- Unsupervised Feature Extraction based on Neuroal Networks or Manifold Learning
- Pronunciation Clustering for Unsupervised Learning of pronunciation dictionaries
- Unsupervised Training of a Grapheme-to-Phoneme component to avoid pronunciation dictionaries