Uncertainty of Phone Voicing and Its Impact on Speech Synthesis

Abstract

While unit selection speech synthesis is not at the centre of research nowadays, it shows its strengths in deployments where fast fixes and tuning possibilities are required. The key part of this method is target and concatenation costs, usually consisting of features manually designed. When there is a flaw in a feature design, the selection may behave in an unexpected way, not necessarily causing a bad quality speech output. One of such features in our systems was the requirement on the match between expected and real units voicing. Due to the flexibility of the method, we were able to narrow the behaviour of the selection algorithm without worsening the quality of synthesised speech.