Singing synthetic method for tone conversion

Abstract

The invention relates to a singing synthetic method for tone conversion. The method comprises the steps: calculating the frequency spectrum features (shown in the description), a fundamental tone frequency (shown in the description) and an aperiodic component (shown in the description) of a singing audio without makeup and acting of a singer through a STRAIGHT algorithm; constructing a singing tone conversion model, wherein the model is generated through a variational self-encoding-generative confrontation network; training the model through the spectrum features of the singer's singing data without makeup and acting, and obtaining a trained singing tone conversion model; finally inputting the spectrum of a to-be-converted singer's singing audio without makeup and acting into the trained singing tone conversion model, wherein the model output is the spectrum with the tone of a target singer (shown in the description), and synthesizing the converted spectrum (shown in the description),fundamental tone frequency (shown in the description) and aperiodic component (shown in the description) into the singing sound through the STRAIGHT algorithm after tone conversion. The method achieves the many-to-many singing conversion under a non-parallel data set.

Description

technical field [0001] The invention belongs to an audio signal processing method in the field of singing, in particular to a singing voice synthesis method oriented to timbre conversion. Background technique [0002] The American National Institute of Standardization defines timbre as follows, "timbre refers to a certain attribute of sound produced by hearing, and the listener can judge the difference between two sounds presented in the same way with the same pitch and loudness. ". Therefore, the vocal timbre during singing refers to the voice characteristics that people use to identify the specific singer when different singers sing the same song. [0003] The timbre conversion process can be understood as a process of modifying and reconstructing the original audio signal. Similar to voice timbre conversion, the so-called singing voice timbre conversion refers to converting the timbre of the singing voice of the source singer to the timbre of the target singer on the pr...

AI Technical Summary

Problems solved by technology

The so-called "one-to-one" means that the speech timbre conversion model can only convert the timbre of a specific source speaker into the timbre of a specific target speaker. Obviously, the generalization ability of this "one-to-one" model is very high. Poor; the so-called parallel data set means that the training data set used by the timbre conversion model in the training stage is strictly parallel, that is, the voice data of the source speaker and the target speaker have exactly the same semantic content. Said that it is very difficult to obtain such a parallel corpus due to the limitation of the singer's own singing repertoire

Therefore, the commonly used "one-to-one" speech timbre conversion model based on parallel corpora is not suitable for singing timbre conversion

Images