Combining monaural source separation with Long Short-Term Memory for increased robustness in vocalist gender recognition