High accurate model-integration-based voice conversion using dynamic features and model structure optimization