An End-to-end Singing Voice Synthesis Method Based On Transinger

Singing voice synthesis is a kind of derivative technology of speech synthesis.It is a classical application of computer technology in the field of music to synthesize electronic singing by extracting information from music score.From the early sample concatenation method,the parametric model to the current neural network model,the performance of the singing voice synthesis system has reached the level that people can accept.At present,the related work is based on the serial model structure,using the lyrics,pitch,beat and other information in the score to train multiple models to work together.In the process of building a singing voice synthesis system,it usually requires many professionals to work together.The model training is difficult and the project time is long.Moreover,the series model is prone to error accumulation,and the final synthesis effect is not stable.Inspired by the related work of speech synthesis,this study proposes a complete end-to-end speech synthesis front-end model,Transinger,to simplify the model construction and training process.According to the characteristics of the front-end model and singing data,the existing neural network vocoder model is studied and improved,the characteristics of acoustic characteristics and voice data are analyzed,and the front-end model is compared Finally,a set of end-to-end voice synthesis method is proposed.Experimental results show that the proposed method can synthesize high quality and natural singing,and the pitch condition control method can enhance the robustness and convergence speed of the vocoder model.In addition,this study also attempts to synthesize singing voice of multiple singers.It is found that the Transinger model can synthesize singing voice of multiple singers at the same time under the condition of adding singer identity information.