Statistical voice conversion using GA-based informative feature

In order to make voice conversion (VC) robust to noise, we propose VC using GA-based informative feature (GIF), by adding an extraction process of GIF to a conventional VC. GIF is proposed as a feature that can be applied not only in pattern recognition but also in relative tasks. In speech recognition, furthermore, GIF could improve recognition accuracy in noise environment. We evaluated the performances of VC using spectral segmental features (conventional method) and GIF, respectively. Objective experimental result indicates that in noise environments, the proposed method was better than the conventional method. Subjective experiment was also conducted to compare the performances. These results show that application of GIF to VC was effective.