In human-machine speech communication, it is challenging to make the machine talk as naturally as human to facilitate "frictionless" human-machine interactions. In this paper, we introduce a "trajectory tiling" based approach to high quality speech rendering, where speech parameter trajectories, extracted from natural, processed, or synthesized speech, are used to guide the search for the best sequence of waveform "tiles" stored in a pre-recorded speech database. The good "tile" candidates at phone, state or frame are then used to construct a lattice-like "sausage". In the sausage, the best path of concatenated tiles is then searched via the Viterbi algorithm. The best string of concatenated waveform segments (tiles) is output as the final rendered speech. The proposed trajectory tiling approach to speech rendering has been tested in three tasks: TTS synthesis, cross-lingual voice transformation for personalized speech-to-speech translation and mixed-code TTS synthesis. Experimental results show that the trajectory tiling approach can yield speech which is natural and highly intelligible. The perceived high quality of rendered speech is also confirmed in both objective and subjective evaluations.
展开▼
