Phase-mismatch-free and data efficient approach to natural sounding harmonic concatenative speech synthesis

This paper proposes our innovative approach to speech signal representation and generation in the harmonic/noise based speech synthesis. The problem with harmonic/noise synthesis arises when it is required to achieve the high-quality synthesis on the low-resource devices. It is because of the necessity to manipulate a large speech unit databases and unavailability of a method for an efficient phase data representation. The other implementations with artificial phases (linear, minimal, zeroed, etc.) produce the synthesized speech with unsatisfactory quality. In proposed approach we use phase data derived from real speech signals to reach natural sounding synthesis. We choose, so-called, representative phase vectors, that are only stored to the speech unit database. We reached a dramatic reduction of demands for the database storage space. It corresponds with the rate (number of voiced speech units):(number of voiced speech frames) stored in the database. Our proposed method also ensures the phase coherence in the synthesized speech signal automatically.