• Ali Khairat
  • 25.10.2010
  • 31.01.2011
Automatic Speech Recognition (ASR) systems work very reliably if close-talking microphones are used for speech input. If the distance between speaker and microphone is increased, additive distortions and reverberation hamper recognition in real-world environments.

Usually close-talking recordings are used for the training of Hidden Markov Model (HMM)-based recognizers. Therefore, the utterances to be recognized can be very different from the models if they are captured by distant-talking microphones. Furthermore, the dispersive effect of reverberation on the speech feature vectors limits the performance of traditional HMM-based recognizers in reverberant environments. The goal of model adaptation algorithms is to adapt the models to the target environment using only a few calibration utterances.

A novel approach [Takiguchi, IEICE Trans. Inf. & Syst., 2006] for adaptation of the models to reverberation is based on first-order linear prediction. By adapting the HMMs of the recognizer in each frame, the dispersive character of reverberation can be accounted for.

In this project, the EM algorithm for estimating the linear prediction coefficients shall be implemented in MATLAB. The performance of the adaptation approach is then to be evaluated for different rooms based on digit recognition experiments.