• Armin Sehr
  • 28.07.2010
  • 27.08.2010
Many real-world Automatic Speech Recognition (ASR) systems are expected to work reliably with distant-talking recordings in a wide range of different acoustic environments. For example, meeting transcription, voice control of interactive television sets, and voice control of humanoid robots should work dependably in different rooms.

The goal of this master thesis is to evaluate different reverberation-robust ASR techniques in various acoustic environments, i.e. under various reverberation conditions. In particular, the ASR performance of HMMs trained on matched reverberant data shall be evaluated in different rooms. Then, it shall be compared to the performance of multi-style HMMs and to the performance of the REMOS concept, a generic framework for reverberation-robust ASR that has been developed in the Audio Research Laboratory of the Chair. To this end, an already available digit recognition system can be used.