Improving speech synthesis quality by reducing pitch peaks in the source recordings