Deep neural network based trainable voice source model for synthesis of speech with varying vocal effort.