Training method, device and electronic equipment for speech spectrum generation model

Abstract

The application discloses a training method, device and electronic equipment for a speech spectrum generation model, and relates to the technical fields of speech synthesis and deep learning. The specific implementation plan is: input the first character sequence into the speech spectrum generation model, generate the analog spectrum sequence corresponding to the first character sequence, and obtain the first loss value of the analog spectrum sequence according to the preset loss function; An analog spectrum sequence corresponding to a text sequence is input into an adversarial loss function model to obtain a second loss value of the analog spectral sequence, and the adversarial loss function model is a generated adversarial network model; according to the first loss value and the second loss value, the speech spectrum generation model is trained. The adversarial loss function model can learn a loss function based on the generative adversarial network, and train the speech spectrum generation model in combination with the preset loss function, so as to make the spectrum sequence generated by the speech spectrum generation model clearer.

Description

technical field [0001] The present application relates to the technical field of data processing, in particular to the technical field of speech synthesis and deep learning, and in particular to a training method, device and electronic equipment for a speech spectrum generation model. Background technique [0002] Spectrum generation technology is a very important technology in speech synthesis. Spectrum acts as a bridge to link the input text sequence with the final synthesized audio. [0003] In the prior art, the speech spectrum generation model uses the mean square error (Mean Square Error, MSE) loss function to feed back the error of the generated spectrum, but modeling based on the MSE loss function cannot reflect the nature of the spectrum, resulting in the generated spectrum being very Vague. When the vocoder is trained with the real clear spectrum, inputting the above-mentioned fuzzy spectrum into the vocoder will cause inconsistencies in the training and judgment ...

AI Technical Summary

Problems solved by technology

[0003] In the prior art, the speech spectrum generation model uses the mean square error (Mean Square Error, MSE) loss function to feed back the error of the generated spectrum, but modeling based on the MSE loss function cannot reflect the nature of the spectrum, resulting in the generated spectrum being very Vague

When the vocoder is trained with the real clear spectrum, inputting the above-mentioned fuzzy spectrum into the vocoder will cause inconsistencies in the training and judgment of the vocoder, which will seriously affect the stability of the vocoder and affect the final synthesis audio quality

Images