Speech enhancement: Simulation and quality evaluation

Project Type: Master/Diploma Thesis
Student: Serdarevic Emir


 The goal of speech enhancement is to eliminate the additive noise present in a speech signal and restore its original form. This work presents several efficient methods for speech enhancement, which are implemented in Matlab. The algorithms only use single-channel noisy speech to estimate the noise level. Finally, two noise estimators are chosen and compared. The results of the best speech enhancement methods were compared with the objective and subjective tests together with results of a chip for speech enhancement made by Cortologic. The best results for quality improvement were obtained with the non-linear gain calculation from Ephraim and Malah combined with the noise estimator based on minimum statistics from R. Martin. In a subjective listening test, 68% of 11 listeners voted for this method concerning the best suppression of the noise and 53% concerning the intelligibility.