Voice conversion model training method and device

Abstract

The embodiment of the invention provides a voice conversion model training method and device, and the method comprises the steps: carrying out the feature extraction of a sample audio, and obtaining a Mel spectrum feature tag and a fundamental frequency sequence; inputting the Mel spectrum feature tag into an encoder to obtain a first content vector; inputting the first content vector into a bottleneck layer to obtain a current codebook vector and a second content vector; determining a first loss value based on the first content vector and the current codebook vector; inputting the first content vector into a perceptron layer to obtain an emission probability of each character or blank character corresponding to the first content vector; determining a second loss value based on the transcriptional text tag and the emission probability of the sample audio; inputting the normalized fundamental frequency sequence, the second content vector and the speaker tag of the sample audio into a decoder to obtain predicted Mel spectrum features; determining a third loss value based on the Mel spectrum feature tag and the predicted Mel spectrum feature; and training a voice conversion model by taking minimization of the loss value as a target.

Description

technical field [0001] This description relates to the technical field of speech processing, in particular to a method and device for training a speech conversion model. Background technique [0002] Voice Conversion (Voice Conversion, VC) is a relatively popular research topic in recent years. It is a process of converting a person's voice into another person's timbre on the basis of retaining complete content information. Speech conversion belongs to a technical field of speech synthesis, and since it focuses on the conversion of identity information in speech, it is one of the challenging research problems in speech signal processing. [0003] The current voice conversion technology relies on parallel corpus, that is, voices with the same content information recorded by different people. However, in actual application scenarios, it is difficult for users to record voices with specific content. It is more difficult, which in turn affects the subsequent voice conversion ef...

AI Technical Summary

Problems solved by technology

[0003] The current voice conversion technology relies on parallel corpus, that is, voices with the same content information recorded by different people, but in actual application scenarios, it is difficult for users to record voices with specific content, so the acquisition of parallel corpus in the real environment is difficult. It is more difficult, which will affect the subsequent voice conversion effect

Images