Speech synthesis method and device

Abstract

The application discloses a speech synthesis method and device. In this method, after obtaining the initial speech point of the initial speech, the corresponding initial speech point vector and the prediction time period, according to the time sequence of the initial speech point, the initial speech point vector is input into the pre-trained speech synthesis network, so that the speech synthesis network outputs sequentially Predicted voice point vectors at each time to be tested in the predicted time period; using a preset neural network algorithm to perform calculations on the predicted voice point vectors at each time to be tested in the predicted time period to synthesize predicted voices in the predicted time period. The method reduces the calculation amount in the speech prediction process and improves the speech synthesis efficiency.

Description

technical field [0001] The present application relates to the technical field of speech synthesis, in particular to a speech synthesis method and device. Background technique [0002] The current mainstream technology of speech synthesis is neural network model, such as Wavenet, Wavernn, Simplernn, etc., which can realize the conversion from text to sound, that is, speech synthesis. When the Wavenet model is applied to speech synthesis, whether it is English or Mandarin Chinese, human experts can obtain the best results in the industry compared with traditional parametric or splicing systems in evaluating its naturalness. In other words, Wavenet can convert text information into smooth and fluent speech through layer-by-layer causal convolutional neural networks as output. [0003] The network structure of Wavenet is a causal convolutional network, which usually includes a 40-layer convolutional neural network (4 blocks, each block has a 10-layer neural network). In the wa...

AI Technical Summary

Problems solved by technology

[0004] However, since each predicted voice point of Wavenet is fed back to the network to predict the next voice point, that is, only one voice point can be predicted at a time, and the same amount of calculations must be performed for the output of each voice point. Network calculation (weight calculation, convolution calculation and skip connect calculation), very time-consuming

For sampling points that are far away from the prediction point (or "sampling points with little relevance"), the neural network calculation of the same amount of calculation is also performed, resulting in a slower prediction speed

Images