Spectral mapping for voice conversion using speaker selection and vector field smoothing