Contributions to Turbo Automatic Speech Recognition

Be it Siri or Amazon Echo - automatic speech recognition is making its way into our lives and despite astonishing improvements in recognition in general, it is still far from being as good as human speech comprehension. In order to open up possible paths for more robust and possibly distributed speech recognition systems, the PhD thesis "Contributions to Turbo Automatic Speech Recognition" deals with a novel method for iterative optimal information fusion.
A fusion is always necessary and profitable when different information sources are to be combined in a statistically optimal way. This can be the combination of audio (speech recognition) and video (lip reading), but also the combination of two similar sensors (two microphones, or for humans the right and left ear). The chosen approach represents the consequent application of the turbo code principle known from communications to questions of automatic speech recognition with multiple data streams. As a major innovation, the PhD thesis presents a so-called modified Viterbi algorithm, which provides a novel information representation for iterative feedback. Two individual recognizers repeatedly evaluate their respective input signal of the underlying speech utterance and exchange information from iteration to iteration, thus moving step by step towards a jointly improved recognition result.