Research of Pitch Controllability Based on LPCNet

With the development of speech synthesis technology in most scenarios, modifying the pitch of audio has been widely used. Currently most pitch controllable methods are achieved by using a separate reference encoder neural network, but this approach requires training more complicated neural networks and does not generalize to unseen speakers without speaker adaptation. To address this problem, a method based on LPCNet vocoder for explicitly pitch control is proposed. Firstly, during the training phase, by optimizing the training features, the model can synthesize more natural and robust voices. Next, the fundamental frequency is adjusted by inputting a control curve in the inference stage, and then the acoustic features are inferred into waveform using the vocoder LPCNet. The experimental results show that the proposed method can control the pitch flexibly and the sound quality of the synthetic voice is improved.