• Christian Kuschel
  • 01.05.2011
  • 31.10.2011
Nowadays, distant-talking microphone systems are employed in various applications, such as hands-free conference telephones and hearing aids. In these scenarios however, the microphones capture, in addition to the desired signal, its reverberation, which is caused by multiple reflections of the sound waves at walls and other objects. Especially in highly reverberant enclosures, this often leads to a significant decrease of the speech intelligibility.

One way to cope with this problem is the application of dereverberation algorithms. They can be divided into different categories according to whether a single or multiple microphones are used. Moreover, such methods can either operate "blindly" or assume prior knowledge.

In this thesis, a single-channel dereverberation algorithm shall be realized, which employs an automatic speech recognition (ASR) system for obtaining additional knowledge on the unobservable clean speech signal. The ASR toolkit HTK shall therefore be adapted to set up pairs of hidden Markov models (HMMs) that are trained on reverberant and clean speech data, respectively. For dereverberation, the microphone signal is to be recognized first, using the reverberantly trained HMM. The according clean speech probability density function (pdf) of each time-frequency bin of the recognized utterance can then be extracted from the clean speech HMM. Since the reverberation tail is considered as uncorrelated additive distortion, a suitable single-channel noise-suppression technique, e.g., Wiener filtering, shall be chosen and used for dereverberation while exploiting the clean speech pdf. To this end, the incorporation of the pdf into the filter is to be theoretically investigated and implemented using Matlab. Finally, the performance of the realized algorithm shall be evaluated based on the TI digit corpus.