In this thesis, the target source extraction system using the constraint on the source location is proposed for speech enhancement.
The generalized sidelobe canceller(GSC) is considered as the most feasible conventional system because of its simplicity and capability of source extraction. But it has the trade-off between the sound quality and the noise reduction capacity. In case of the high learning rate, although noises are reduced well, the sound quality is good in the low learning rate. The blind source separation(BSS) is also used to extract the target signal. because the independent component component analysis (ICA)-based BSS is effective. But, It is impractical in real environments because of its clustering technique. Also, it is poor in various noise environments including music noises.
The proposed system uses the TF masking technique which is one of the BSS methods. Our proposed TF masking technique compensates the shortage of the conventional TF masking technique. It becomes useful and feasible under the constraint on the source location. And our proposed system can run on-line and is possible to extract only the frontal target source corrupted by noise sources, pretty well.
The performance of our proposed technique was evaluated by the perceptual evaluation of speech quality (PESQ) score for noisy sentences recorded in real environment. We confirmed that the proposed system increased the PESQ score as 0.4 compared to the conventional system which increased as 0.06 and 0.1 in speech and music noise environments, respectively.