Speech synthesis method based on generative adversarial network

Abstract

The invention discloses a speech synthesis method based on a generative adversarial network. The method comprises the following steps of: S1, preparing training samples, including real audio data, and extracting Mel spectrum features; S2, setting an initialized generator parameter set and a discriminator parameter set; S3, inputting the Mel spectrum features to a generator to obtain an output synthetic audio; S4, correspondingly and simultaneously inputting the real audio data in the S1 and the output synthesized audio obtained in the S3 into a multi-dimensional discriminator and a pooling discriminator; S4, inputting output results of the discriminators into a loss function formula, and respectively calculating a generator loss function and a discriminator loss function; S5, updating the generator and the discriminators; S6, repeating by using the updated generator and discriminators until the set maximum value M of the number of updating times is reached; S7, after each update, returning to the step S3; and S8, performing speech synthesis by using the generator. According to the method, one-dimensional convolution operation is adopted by the generator, model parameters are small, and the speed is high.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence speech synthesis, and in particular relates to a speech synthesis method based on a generative confrontation network. Background technique [0002] Voice, as the most direct and fast way of communication, plays a very important role in the field of artificial intelligence, and has been widely used in robots, cars, synthetic anchors and other fields. With the wide application of artificial intelligence products, the requirements for the naturalness, clarity, and intelligibility of speech synthesis are getting higher and higher. Deep learning has enabled the rapid development of speech synthesis technology. [0003] The commonly used deep learning-based speech synthesis scheme is mainly divided into two stages: predicting its acoustic features based on text information, such as mel-spectrograms; predicting its original audio waveform from acoustic features, that is, vocoder model le...

AI Technical Summary

Problems solved by technology

WaveNet is an autoregressive convolutional neural network. As the first batch of deep learning algorithms used for speech synthesis, it greatly improves the quality of speech synthesis. However, its model structure determines that the speed is very slow and it is difficult to apply it to actual products.

In recent years, the research on the speech synthesis vocoder has mainly focused on improving the calculation speed and reducing the model parameters, and there is no significant improvement in the synthesis speed

Images