Transforming Pitch in a Voice Conversion Framework

A subtask of voice conversion is to accurately map the pitch contour of a source speaker to a target speaker. So far, the most widely employed method for carrying out this mapping is based on adjusting the pitch range of the source speaker to match the target while keeping the shape of the contour unchanged. In this project, we investigate four alternative algorithms for pitch contour mapping and compare their performance with the popular baseline method in an objective framework as well as through perceptual tests. The first two methods extend the baseline to allow more complex mappings of the pitch range, while the last two methods aim to impart an entirely new contour onto the target. We have found that all four methods improve the baseline for most cases. The amount of improvement, however, varies from method to method and manifests a dependency on the nature of the training data available.